scale-cv-dl

Why Deep Learning Has Not Superseded Traditional Computer Vision

Posted on by Zbigatron

This is another post that’s been inspired by a question that has been regularly popping up in forums:

Has deep learning superseded traditional computer vision?

Or in a similar vein:

Is there still a need to study traditional computer vision techniques when deep learning seems to be so effective? 

These are good questions. Deep learning (DL) has certainly revolutionised computer vision (CV) and artificial intelligence in general. So many problems that once seemed improbable to be solved are solved to a point where machines are obtaining better results than humans. Image classification is probably the prime example of this. Indeed, deep learning is responsible for placing CV on the map in the industry, as I’ve discussed in previous posts of mine.

But deep learning is still only a tool of computer vision. And it certainly is not the panacea for all problems. So, in this post I would like to elaborate on this. That is, I would like to lay down my arguments for why traditional computer vision techniques are still very much useful and therefore should be learnt and taught.

I will break the post up into the following sections/arguments:

  • Deep learning needs big data
  • Deep learning is sometimes overkill
  • Traditional CV will help you with deep learning

But before I jump into these arguments, I think it’s necessary to first explain in detail what I mean by “traditional computer vision”, what deep learning is, and also why it has been so revolutionary.

Background Knowledge

Before the emergence of deep learning if you had a task such as image classification, you would perform a step called feature extraction. Features are small “interesting”, descriptive or informative patches in images. You would look for these by employing a combination of what I am calling in this post traditional computer vision techniques, which include things like edge detection, corner detection, object detection, and the like.

In using these techniques – for example, with respect to feature extraction and image classification – the idea is to extract as many features from images of one class of object (e.g. chairs, horses, etc.) and treat these features as a sort of “definition” (known as a bag-of-words) of the object. You would then search for these “definitions” in other images. If a significant number of features from one bag-of-words are located in another image, the image is classified as containing that specific object (i.e. chair, horse, etc.).

The difficulty with this approach of feature extraction in image classification is that you have to choose which features to look for in each given image. This becomes cumbersome and pretty much impossible when the number of classes you are trying to classify for starts to grow past, say, 10 or 20. Do you look for corners? edges? texture information? Different classes of objects are better described with different types of features. If you choose to use many features, you have to deal with a plethora of parameters, all of which have to be fine-tuned by you.

Well, deep learning introduced the concept of end-to-end learning where (in a nutshell) the machine is told to learn what to look for with respect to each specific class of object. It works out the most descriptive and salient features for each object. In other words, neural networks are told to discover the underlying patterns in classes of images.

So, with end-to-end learning you no longer have to manually decide which traditional computer vision techniques to use to describe your features. The machine works this all out for you. Wired magazine puts it this way:

If you want to teach a [deep] neural network to recognize a cat, for instance, you don’t tell it to look for whiskers, ears, fur, and eyes. You simply show it thousands and thousands of photos of cats, and eventually it works things out. If it keeps misclassifying foxes as cats, you don’t rewrite the code. You just keep coaching it.

The image below portrays this difference between feature extraction (using traditional CV) and end-to-end learning:

traditional-cv-and-dl

So, that’s the background. Let’s jump into the arguments as to why traditional computer vision is still necessary and beneficial to learn.

Deep Learning Needs Big Data

First of all, deep learning needs data. Lots and lots of data. Those famous image classification models mentioned above are trained on huge datasets. The top three of these datasets used for training are:

Easier tasks than general image classification will not require this much data but you will still need a lot of it. What happens if you can’t get that much data? You’ll have to train on what you have (yes, some techniques exist to boost your training data but these are artificial methods).

But chances are a poorly trained model will perform badly outside of your training data because a machine doesn’t have insight into a problem – it can’t generalise for a task without seeing data.

And it’s too difficult for you to look inside the trained model and tweak things around manually because a deep learning model has millions of parameters inside of it – each of which is tuned during training. In a way, a deep learning model is a black box.

Traditional computer vision gives you full transparency and allows you to better gauge and judge whether your solution will work outside of a training environment. You have insight into a problem that you can transfer into your algorithm. And if anything fails, you can much more easily work out what needs to be tweaked and where.

Deep Learning is Sometimes Overkill

This is probably my favourite reason for supporting the study of traditional computer vision techniques.

Training a deep neural network takes a very long time. You need dedicated hardware (high-powered GPUs, for example) to train the latest state-of-the-art image classification models in under a day. Want to train it on your standard laptop? Go on a holiday for a week and chances are the training won’t even be done when you return.

Moreover, what happens if your trained model isn’t performing well? You have to go back and redo the whole thing again with different training parameters. And this process can be repeated sometimes hundreds of times.

But there are times when all this is totally unnecessary. Because sometimes traditional CV techniques can solve a problem much more efficiently and in fewer lines of code than DL. For example, I once worked on a project to detect if each tin passing through on a conveyor belt had a red spoon in it. Now, you can train a deep neural network to detect spoons and go through the time-consuming process outlined above, or you can write a simple colour thresholding algorithm on the colour red (any pixel within a certain range of red is coloured white, every other pixel is coloured black) and then count how many white pixels you have. Simple. You’re done in an hour!

Knowing traditional computer vision can potentially save you a lot of time and unnecessary headaches.

Traditional Computer Vision will Improve your Deep Learning Skills

Understanding traditional computer vision can actually help you be better at deep learning.

For example, the most common neural network used in computer vision is the Convolutional Neural Network. But what is a convolution? It’s in fact a widely used image processing technique (e.g. see Sobel edge detection). Knowing this can help you understand what your neural network is doing under the hood and hence design and fine-tune it better to the task you’re trying to solve.

Then there is also a thing called pre-processing. This is something frequently done on the data that you’re feeding into your model to prepare it for training. These pre-processing steps are predominantly performed with traditional computer vision techniques. For example, if you don’t have enough training data, you can do a task called data augmentation. Data augmentation can involve performing random rotations, shifts, shears, etc. on the images in your training set to create “new” images. By performing these computer vision operations you can greatly increase the amount of training data that you have.

Conclusion

In this post I explained why deep learning has not superseded traditional computer vision techniques and hence why the latter should still be studied and taught. Firstly, I looked at the problem of DL frequently requiring lots of data to perform well. Sometimes this is not a possibility and traditional computer vision can be considered as an alternative in these situations. Secondly, occasionally deep learning can be overkill for a specific task. In such tasks, standard computer vision can solve a problem much more efficiently and in fewer lines of code than DL. Thirdly, knowing traditional computer vision can actually make you better at deep learning. This is because you can better understand what is happening under the hood of DL and you can perform certain pre-processing steps that will improve DL results.

In a nutshell,  deep learning is just a tool of computer vision that is certainly not a panacea. Don’t only use it because it’s trendy now. Traditional computer vision techniques are still very much useful and knowing them can save you time and many headaches.

To be informed when new content like this is posted, subscribe to the mailing list:

brain-reading

Machines That Can Read Our Minds from MRI Scans

Posted on by Zbigatron

Who here watches Black Mirror? Oh, I love that show (maybe except for Season 4 – only one episode there was any good, IMHO). For those that haven’t seen it yet, Black Mirror basically tries to extrapolate what society will look like if technological advances were to continue on their current trajectory.

A few of the technologies presented in that show are science fiction on steroids. But I think that many are not and that it’s only a matter of time before we reach that (dark?) high-tech reality.

A reader of this blog approached me recently with a (preprint) academic publication that sounded pretty much like it had been taken out of that series. It was a paper purporting that scientists can now reconstruct our visual thoughts by analysing our brain scans.

“No way. Too good to be true”, was my first reaction. There has to be a catch. So, I went off to investigate further.

And boy was my mind blown. Allow me in this post to present to you that publication out of Kyoto, Japan. Try telling me at the end that you’re not impressed either!

(Note: Although this post is technically not directly related to computer vision, I’m still including it in my blog because it deals with trying to understand how our brains work when it comes to visual stimuli. This is important to do because many times in the history of computer vision and AI in fact, it has been shown that biologically inspired solutions (i.e. solutions that mimic the way we behave) can obtain really good if not superior results to other solutions. The prime example of this, of course, is neural networks)

Machines that can read our minds

The paper in question is entitled “Deep image reconstruction from human brain activity” (Shen et al., bioRxiv, 2017). Although it has been published in preprint, meaning that it hasn’t been peer-reviewed and technically doesn’t hold too much academic weight, the site that it was published on is of repute (www.biorxiv.org) as are the institutes that co-authored the work (ATR Computational Neuroscience Laboratories and Kyoto University, the latter being one of Asia’s highest ranked universities). Hence, I’m confident in the credibility of their research.

In a nutshell, the paper is claiming to have developed a way to reconstruct images that people are looking at or thinking about by solely analysing functional magnetic resonance imaging (fMRI). fMRI machines measure brain activity by detecting changes in blood flow through regions of the brain – generally speaking, the more blood flow, the more brain activity. Here’s an example fMRI image showing increased brain activity in orange/yellow:

(image source)

Let’s look at some details of their paper.

The first step was to utilise machine learning to examine what the relationship is between brain activity as detected by fMRI scans and the images the were being shown to test subjects. What’s interesting is that there was a focus on mapping hierarchical features between the fMRI images and real images.

Hierarchical features is a way to deconstruct images into an ascending level of detail. For example, to describe a car you could start by describing edges, then working up to curves, and then particular shapes like circles, and then objects like wheels; and then finally you’d get four wheels, a chassis and some windows.

Deconstructing images by looking at hierarchical features (edges, curves, shapes, and so on) is also the way neural networks and our brains work with respect to visual stimuli. So, this plan of action to focus on hierarchical features in fMRI images seems intuitive.

The hierarchical features extracted from fMRI images were then used to replace the features of a deep neural network (DNN). Next, an iterative process was started where base images (produced by a deep generator network) were fed into the DNN and at each iteration individual pixels of the image were transformed until the pixel values matched the features of the DNN (which were extracted from fMRI images) to a certain threshold (error) level.

It’s a complicated process, I know. I’ve tried to summarise it as best as I could here! To understand it more, you’ll need to go back to the original publication and then follow a trail of referenced articles. The image below shows this process – not that it helps much 🙂 But I think the most important part of this study are the results – that’s coming up in the next section.

deep-image-reconstruction
Deep image reconstruction (source: original publication)

Test subjects were shown various images from three different classes: natural colour images (sampled from ImageNet), artificial geometrical shapes (in 8 colours), and black and white alphabetical letters. Scans of brain activity were conducted during the viewing sessions.

Interestingly, test subjects were also asked to recollect images they had been previously shown and brain scans during this task were also taken. The idea was to try to see if a machine could truly discern what a person was actually thinking rather than just viewing.

Results from the Study

This is where the fun begins!

But before we dive into the results, I need to mention that this is not the first time fMRI imaging has been used to attempt to reconstruct images. Similar projects (referenced in the publication we’re discussing) go back to 2011 and 2013. The novelty of this study, however, is to work with hierarchical features rather than working with fMRI images directly.

Take a look at this video showing past results from previous studies. The images are monochromatic and of minimal resolution:

Now let’s take a look at some results from this particular study. (All images are taken from the original publication.)

How about a reconstructed image of a swan? Remember, this reconstructed image (right) is generated from someone’s brain scan (taken while the test subject was viewing the original image):

Example-Result-Swan
Example result #1: the test subject was shown the image on the left; the image on the right is the reconstructed image from brain scans.

You must be impressed with that!

Here’s a reconstructed picture of a duck:

experiment-result-duck
Example result #2: the test subject was shown the image on the left; the image on the right is the reconstructed image from brain scans.

For a presentation of results, take a look at this video released by the authors of the paper:

What about those letters I mentioned that were also shown to test subjects? Check this out:

results-from-text-experiment
Example results #3: the test subject was shown the letters in the top row; the bottom row shows the reconstructed images from brain scans.

Scary how good the reconstructions are, isn’t it? Imagine being able to tell what a person is reading at a given time!? I’ve already written a post about non-invasive lie detection from thermal imaging. If we can just work out how to get fMRI images like that (i.e. without engaging the subject), we’ll be able to tell what a person is looking at without them even knowing about it. Black Mirror stuff!

Results from the artificial geometric shapes images can be viewed here – also impressive, of course.

And how about reconstructed images from scans performed when people were thinking about the images they had been previously shown? Well, results here aren’t as impressive (see below). The authors concede this also. But hey! One step at a time, right?

results-imagined-images
Example results #4: the test subject was told to think about the images in the rop row; the bottom row shows the attempted reconstructions of these images from brain scans.

Conclusion

In this post I presented a preprint publication in which it is purported that scientists were able to reconstruct visual thoughts by analysing brain scans. fMRI images were analysed for hierarchical features that were then used in a deep neural network. Base images were fed through this network and iteratively transformed until the features of each image matched the features in the neural network. Results from this study were presented in this post and I think we can all agree that they were quite impressive. This is the future, folks! Black Mirror stuff perhaps.

The uses for good of this technology are numerous – communicating with people in comas is one such use case. Of course, the abuse that could come from this is frightening, too. Issues with privacy (ah, that pesky perennial question in computer science!) spring to mind immediately.

But as you may have gathered from all my posts here, I look forward to what the future holds for us, especially with respect to artificial intelligence. And likewise with this technology. I can’t wait to see it progress.

To be informed when new content like this is posted, subscribe to the mailing list:

Amazon Go – Computer Vision at the Forefront of Innovation

Posted on by Zbigatron

Where would a computer vision blog be without a post about the new cashier-less store recently opened to the public by Amazon? Absolutely nowhere.

But I don’t need additional motivation to write about Amazon Go (as the store is called) because I am, to put it simply, thrilled and excited about this new venture. This is innovation at its finest where computer vision is playing a central role.

How can you not get enthusiastic about it, then? I always love it when computer vision makes the news and this is no exception.

In this post, I wish to talk about Amazon Go under four headings:

  1. How it all works from a technical (as much as is possible) and non-technical perspective,
  2. Some of the reported issues prior to public opening,
  3. Some reported in-store issues post public opening, and
  4. Some potential unfavourable implications cashier-less stores may have in the future (just to dampen the mood a little)

So, without further ado…

How it Works – Non-Technically & Technically

The store has a capacity of around 90 people – so it’s fairly small in size, like a convenience store. To enter it you first need to download the official Amazon app and connect it to your Amazon Prime account. You then walk up to a gate like you would at a metro/subway and scan a QR code from the app. The gate opens and your shopping experience begins.

Inside the store, if you wish to purchase something, you simply pick it up off the shelf and put it in your bag or pocket. Once you’re done, you walk out of the shop and a few minutes later you get emailed a receipt listing all your purchases. No cashiers. No digging around for money or cards. Easy as pie!

What happens on the technical side of things behind the scenes? Unfortunately, Amazon hasn’t disclosed much at all, which is a bit of a shame for nerds like me. But I shouldn’t complain too much, I guess.

What we do know is that sensor fusion is employed (sensor fusion is when data is combined from multiple sensors/sources to provide a higher degree of accuracy) along with deep learning.

Hundreds of cameras and depth sensors are attached to the ceiling around the store:

amazon-go-cameras
The cameras and depth sensors located on the ceiling of the store (image source)

These track you and your movements (using computer vision!) throughout your expedition. Weight sensors can also be found on the shelves to assist the cameras in discerning which products you have chosen to put into your shopping basket.

(Note: sensor fusion is also being employed in autonomous cars. Hopefully I’ll be writing about this as well soon.)

In 2015, Amazon filed a patent application for its cashier-less store in which it stated the use of RGB cameras (i.e. colour cameras) along with facial recognition. TechCrunch, however, has reported that the Vice President of Technology at Amazon Go told them that no facial recognition algorithms are currently being used.

In-Store Issues Prior to Public Opening

Although the store opened its doors to the public a few weeks ago, it has been open to employees since December 2016. Initially, Amazon expected the store to be ready for public use a few months after that but public opening was delayed by nearly a year due to “technical problems“.

We know what some of the dilemmas behind these “technical problems” were.

Firstly, Amazon had problems tracking more than 20 people in the store. If you’ve ever worked on person-tracking software, you’ll know how hard it is to track a crowd of people with similar body types and wearing similar clothes. But it looks like this has been resolved (to at least a satisfactory level for them). It’s a shame for us, though, to not be given more information on how Amazon managed to get this to work.

Funnily enough, some employees of Amazon knew about this problem and in November last year tried to see if a solution had been developed. Three employees dressed up in Pikachu costumes (as reported by Bloomberg here) while doing their round of shopping to attempt to fool the system. Amazon Go passed this thorough, systematic, and very scientific test. Too bad I couldn’t find any images or videos of this escapade!

We also know that initially engineers were assisting the computer vision system behind the scenes. The system would let these people know when it was losing confidence with its tracking results and would ask them to intervene, at least minimally. Nobody is supposedly doing this task any more.

Lastly, I also found information stating that the system would run into trouble when products were taken off the shelf and placed back on a different shelf. This was reported to have occurred when employees brought their children into the store and they ran wild a little (as children do).

This also appears to have been taken care of because someone from the public attempted to do this on purpose last week (see this video) but to no adverse effects, it would seem.

It’s interesting to see the growing pains that Amazon Go had to go through, isn’t it? How they needed an extra year to try to iron out all these creases. This is such a huge innovation. Makes you wonder what “creases” autonomous cars will have when they become more prominent!

In-Store Issues Post Public Opening

But, alas. It appears as though not all creases were ironed out to perfection. Since Amazon Go’s opening a few weeks ago, two issues have been written about.

The first is of Deirdre Bosa of CNBC not being charged for a small tub of yoghurt:


The Vice President of Amazon Go responded in the following way:

First and foremost, enjoy the yogurt on us. It happens so rarely that we didn’t even bother building in a feature for customers to tell us it happened. So thanks for being honest and telling us. I’ve been doing this a year and I have yet to get an error.

The yoghurt manufacturer replied to that tweet also:

To which Dierdre responded: “Thanks Siggi’s! But I think it’s on Amazon :)”

LOL! 🙂

But as Amazon Go stated, it’s a rarity for these mistakes to happen. Or is that only the case until someone works out a flaw in the system?

Well, it seems as though someone has!

In this video, Tim Pool states that he managed to walk out of the Amazon Go store with a bag full of products and was only charged for one item. According to him it is “absurdly easy to take a bag full of things and not get charged”. That’s a little disconcerting. It’s one thing when the system makes a mistake every now and then. It’s another thing when someone has worked out how to break it entirely.

Tim Pool says he has contacted Amazon Go to let them know of the major flaw. Amazon confirmed with him that he did not commit a crime but “if used in practice what we did would in fact be shoplifting”.

Ouch. I bet engineers are working on this frantically as we speak.

One more issue worth mentioning that isn’t really a flaw but could also be abused is that at the moment you can request a refund on any item without returns. No questions asked. Linus Tech Tips shows in this video how easily this can be done. Of course, since your Amazon Go account needs to be linked to your Amazon Prime account, if you do this too many times, Amazon will catch on and will probably take some form of preventative action against you or will even verify everything by looking back at past footage of you.

Cons of Amazon Go

Like I said earlier, I am really excited about Amazon Go. I always love it when computer vision spearheads innovation. But I also think it’s important to in this post also talk about potential unfavourable implications of a cashier-less store.

Potential Job Losses

The first most obvious potential con of Amazon Go is the job losses that might ensue if this innovation catches on. Considering that 3.5 million people in the US are employed as cashiers (it’s the second-most common job in that country), this issue needs to be raised and discussed. Heck, there have already been protests in this respect outside of Amazon Go:

amazon-go-protest
Protests in front of the Amazon Go store (image source)

Bill Ingram, the organiser of the protest shown above asks: “What will all the cashiers do once their jobs are automated?”

Amazon, not surprisingly, has issued statements on this topic. It has said that although some jobs may be taken by automation, people can be relocated to improve other areas of the store by, for example:

Working in the kitchen and the store, prepping ingredients, making breakfast, lunch and dinner items, greeting customers at the door, stocking shelves and helping customers

Let’s also not forget that new jobs have also been created. For example, additional people need to be hired to manage the technological infrastructure behind this huge endeavour.

Personally, I’m not a pessimist about automation either. The industrial revolution that brought automation to so many walks of life was hard at first but society found ways to re-educate into other areas. The same will happen, I believe, if cashier-less stores become a prominent thing (and autonomous cars also, for that matter).

An Increase in Unhealthy Impulse Purchases

Manoj Thomas, a professor of marketing at Cornell University, has stated that our shopping behaviour will change around cashier-less stores:

[W]e know that when people use any abstract form of payment, they spend more. And the type of products they choose changes too.

What he’s saying is that psychological research has shown that the more distance we put between us and the “pain of paying” the more discipline we need to avoid those pesky impulse purchases. Having cash physically in your hand means you can what you’re doing with your money more easily. And that extra bit of time waiting in line at the cashier could be time enough to reconsider purchasing that chocolate and vanilla tub of ice cream :/

Even More Surveillance

And then we have the perennial question of surveillance. When is too much, too much? How much more data about us can be collected?

With such sophisticated surveillance in-store, companies are going to have access to even more behavioural data about us: which products I looked at for a long time; which products I picked up but put back on the shelf; my usual path around a store; which advertisements made me smile – the list goes on. Targeted advertising will become even more effective.

Indeed, Bill Ingram’s protest pictured above was also about this (hence why masks were worn to it). According to him, we’re heading in the wrong direction:

If people like that future, I guess they can jump into it. But to me, it seems pretty bleak.

Harsh, but there might be something to it.

Less Human Interaction

Albert Borgmann, a great philosopher on technology, coined the term device paradigm in his book “Technology and the Character of Contemporary Life” (1984). In a nutshell, the term is used to explain the hidden, detrimental nature and power of technology in our world (for a more in-depth explanation of the device paradigm, I highly recommend you read his philosophical works).

One of the things he laments is how we are increasingly losing daily human interactions due to the proliferation of technology. The sense of a community with the people around us is diminishing. Cashier-less stores are pushing this agenda further, it would seem. And considering, according to Aristotle anyway, that we are social creatures, the more we move away from human interaction, the more we act against our nature.

The Chicago Tribune wrote a little about this at the bottom of this article.

Is this something worth considering? Yes, definitely. But only in the bigger picture of things, I would say. At the moment, I don’t think accusing Amazon Go of trying to damage our human nature is the way to go.

Personally, I think this initiative is something to celebrate – albeit, perhaps, with just the faintest touch of reservation. 

Summary

In this post I discussed the cashier-less store “Amazon Go” recently opened to the public. I looked at how the store works from a technical and non-technical point of view. Unfortunately, I couldn’t say much from a technical angle because of the little amount of information that has been disclosed to us by Amazon. I also discussed some of the issues that the store has dealt with and is dealing with now. I mentioned, for example, that initially there were problems in trying to track more than 20 people in the store. But this appears to have been solved to a satisfactory level (for Amazon, at least). Finally, I dampened the mood a little by holding a discussion on the potential unfavourable implications that a proliferation of cashier-less stores may have on our societies. Some of the issues raised here are important but ultimately, in my humble opinion, this endeavour is something to celebrate – especially since computer vision is playing such a prominent role in it.

To be informed when new content like this is posted, subscribe to the mailing list:

graph-upward-trend-why

The Reasons Behind the Recent Growth of Computer Vision

Posted on by Zbigatron

In my previous post I looked at the unprecedented growth of computer vision in the industry. 10 years ago computer vision was nowhere to be seen outside of academia. But things have since changed significantly. A telling sign of this is the consistent tripling each year of venture capital funding in computer vision. And Intel’s acquisition of Mobileye in March 2017 for a whopping US$15.3 billion just sums up the field’s achievements.

In that post, however, I only briefly touched upon the reasons behind this incredible growth. The purpose of this article, then, is to fill that gap.

In this respect, I will discuss the top 4 reasons behind the growth of computer vision in the industry. I’ll do so in the following order:

  1. Advancements in hardware
  2. The emergence of deep learning
  3. The advent of large datasets
  4. The increase in computer vision applications

Better and More Dedicated Hardware

I mentioned in another post of mine that one of the main reasons why image processing is such a difficult problem is that it deals with an immense amount of data. To process this data you need memory and processing power. These have been increasing in size and power regularly for over 50 years (cf. Moore’s Law).

Such increases have allowed for algorithms to run faster to the point where more and more things are now capable of being run in real-time (e.g. face recognition).

We have also seen an emergence and proliferation of dedicated pieces of hardware for graphics and image processing calculations. GPUs are the prime example of this. A GPU clock speed may generally be slower than a regular CPU’s, but it can still outperform one for these specific tasks. 

Dedicated pieces of hardware are becoming so highly prized nowadays in computer vision that numerous companies have started designing and producing them. Just two weeks ago, for example, Ambarella announced two new chips designed for computer vision processing with chiefly autonomous cars, drones, and security cameras in mind. And last year, Mythic, a startup based in California, raised over US$10 million to commercialise their own deep learning focused pieces of hardware.

The Emergence of Deep Learning

Deep learning, a subfield of machine learning, has been revolutionary in computer vision. Because of it machines are now getting better results than humans in important tasks such as image classification (i.e. detecting what object is in an image).

Previously, if you had a task such as image classification, you would perform a step called feature extraction. Features are small “interesting”, descriptive or informative patches in images. The idea is to extract as many of these from images of one class of object (e.g. chairs, horses, etc.) and treat these features as a sort of “definition” (known as a bag-of-words) of the object. You would then search for these “definitions” in other images. If a significant number of features from one bag-of-words are located in another image, the image is classified as containing that specific object (i.e. chair, horse, etc.).

The difficulty with this approach is that you have to choose which features to look for in each given image. This becomes cumbersome and pretty much impossible when the number of classes you are trying to classify for starts to grow past, say, 10 or 20. Do you look for corners? edges? texture information? Different classes of objects are better described with different types of features. If you choose to use many features, you have to deal with a plethora of parameters, all of which have to be fine-tuned.

Well, deep learning introduced the concept of end-to-end learning where (in a nutshell) the machine is told to learn what to look for with respect to each specific class of object. It works out the most descriptive and salient features for each object. In other words, neural networks are told to discover the underlying patterns in classes of images.

The image below portrays this difference between feature extraction and end-to-end learning:

traditional-cv-and-dl

Deep learning has proven to be extremely successful for computer vision. If you look below at the graph I used in my previous post showing capital investments into US-based computer vision companies since 2011, you can see that when deep learning became mainstream in around 2014/2015, investments suddenly doubled and have been growing at a regular rate since.

cv-growth-graph
(image source)

You can safely say that deep learning put computer vision on the map in the industry. Without it, chances are we would all still be stuck with it in academia (not that there’s anything wrong with academia, of course).

Large Datasets

To allow a machine to learn the underlying patterns of classes of objects it needs A LOT of data. That is, it needs large datasets. More and more of these have been emerging and have been instrumental in the success of deep learning and therefore computer vision.

Before around 2012, a dataset was considered relatively large if it contained 100+ images or videos. Now, datasets exist with numbers ranging in the millions.

Here are some of the most known image classification databases currently being used to test and train the latest state-of-the-art object classification/recognition models. They have all been meticulously hand annotated by the open source community.

  • ImageNet – 15 million images, 22,000 object categories. It’s HUGE! (I hope to write more about this dataset in the near future, so stay tuned for that).
  • Open Images – 9 million images, 5,000 object categories.
  • Microsoft Common Objects in Context (COCO) – 330K images, 80 object categories.
  • PASCAL VOC Dataset – a few versions exist, 20 object categories.
  • CALTECH-101 – 9,000 images with 101 object categories.

I need to also mention Kaggle, the University of California, Irvine Machine Learning Repository. Kaggle hosts 351 image datasets ranging from flowers, wines, forest fires, etc. It is also the home of the famous Facial Expression Recognition Challenge (FER). The aim of this competition is to correctly detect the emotion of people from seven different categories from nearly 35,000 images of faces.

All these datasets and many more have raised computer vision to its current position in the industry. Certainly, deep learning would not be where it is now without them.

More Applications

Faster machines, larger memories, and other advances in technology have increased the number of useful things machines have been able to do for us in our lives. We now have autonomous cars (well, we’re close to having them), drones, factory robots, cleaning robots – the list goes on. With an increase in such vehicles, devices, tools, appliances, etc. has come an increase in the need for computer vision.

Let’s take a look at some examples of recent new ways computer vision is being used today.

Walmart, for example, a few months ago released shelf-scanning robots into 50 of its warehouses. The purpose of these robots is to detect out-of-stock and missing items and other things such as incorrect labelling. Here’s a picture of one of these robots at work:

(image source)

A British online supermarket is using computer vision to determine the best ways to grasp goods for its packaging robots. This video shows their robot in action:

Agriculture as well is capitalising on the growth of computer vision. iUNU, for example, is developing a network of cameras on rails to assist greenhouse owners to keep track of how their plants are growing.

The famous Roomba autonomous vacuum cleaner got an upgrade a few years ago with a new computer vision system to more smartly manoeuvre around your home.

And our phones? Plenty of computer vision being used in them! Ever noticed your phone camera tracking and focusing on your face when you’re trying to take a picture? That’s computer vision. And how about the face recognition services to unlock your phones? Well, Samsung’s version of it can be classified as computer vision (I write about it in this post).

There’s no need to mention autonomous cars here. We are constantly hearing about them on the news. It’s only a matter of time before we’ll be jumping into one.

Computer vision is definitely here to stay. In fact, it’s only going to get bigger with time.

Summary

In this post I looked at the four main reasons behind the recent growth of computer vision: 1) the advancements in hardware such as faster CPUs and availability of GPUs; 2) the emergence of deep learning, which has changed our way of performing tasks such as image classification; 3) the advent of large datasets that have allowed us to more meticulously study the underlying patterns in images; and 4) the increase in computer vision applications.

All these factors (not always mutually exclusive) have contributed to the unprecedented position of computer vision in the industry.

As I mentioned in my previous post, it’s been an absolute pleasure to have witnessed this growth and to have seen these factors in action. I truly look forward to what the future holds for computer vision.

To be informed when new content like this is posted, subscribe to the mailing list:

graph-chart

The Growth of Computer Vision in the Industry

Posted on by Zbigatron

I started out in computer vision in 2004. I was walking along the corridors of the computer science department at the University of Adelaide (in South Australia) looking at notices put up by lecturers advertising potential undergraduate thesis topics. There wasn’t much there for me until one particular topic caught my eye: developing a vision system for a soccer-playing robot.

Well, the nerd in my awoke! I knocked on the lecturer’s door and 5 minutes later I walked out with a thesis topic and one of those cheeky smiles that said “we’re in for a lot of fun here”. Little did I know that my topic choice was to be the beginning of my adventures in computer vision that would lead me to a PhD and working for companies in Europe and Australia in this field.

So, I’ve been around the world of computer vision for nearly 15 years. When I started out, computer vision was predominantly a research-based field that rarely ventured outside those university corridors and lecture theatres that I used to saunter around. You just couldn’t do anything practical with it – mainly because machines were too slow and memory sizes were too small.

(Note: see this earlier post of mine that discusses why image processing is such a computation and memory demanding activity)

But things have changed since those days. Computer vision has grown immensely and most importantly it’s shaping up to be a viable source of income in the industry. It’s truly been a pleasure to witness this transformation. And it seems as though things are only going to get better.

In this post, then, I would like to present to you how much computer vision in the industry has grown over the last few years and whether this growth will continue in the future. I would also like to briefly talk about what this means for us computer vision enthusiasts in terms of jobs and opportunities in the workforce (where I am based now).

(Note: in my next post I write in more detail about the reasons behind this growth)

Computer vision in the last few years

Recently, growth in computer vision in the industry has surged. To understand just how much, all one has to do is analyse the speed at which top tech corporations are moving into the field now.

Apple, for example, in this respect made at least two significant takeovers last year, both for an undisclosed amount: one of an Israeli-based startup in February 2017 called Realface that works on facial recognition technology for the authentication of users (could it be behind FaceID?); and another in September of 2017 when it acquired Regaind, a startup from Paris that focuses on AI-driven photo and facial analysis.

Facebook has joined the game also. Two months ago (Nov, 2017) it bought out a German computer vision startup called Fayteq. Fayteq develops plugins for various applications that allow you to add or remove objects from existing videos. Its acquisition follows the purchase of Source3, a company that develops video piracy detection algorithms, which also took place in 2017.

While on the topic of social media, let’s take a look at the recent moves made by Twitter and Snapchat. In 2016 Twitter bought Magic Pony Technology for $150 million. Magic Pony Technology employs machine learning to improve low-quality videos on-the-fly by detecting patterns and textures over time. Whereas Snapchat, also in 2016, acquired Seene, which allows you to, among other things, take 3D shots of objects (e.g. 3D selfies) and insert them into videos. Take a look at what Seene can do in this neat little demo video. Apologies for the digression but it’s just too good not to share here:

Amazon has noted the growth of computer vision in the industry so much so that it recently (Oct 2017) created an AI research hub in Germany that focuses on computer vision. This follows shortly upon its acquisition of a 3D body model startup for around $70 million.

The clear stand-out takeover, however, was made by Intel. Last year in March it bought out Mobileye for a WHOPPING $15.3 billion. Mobileye, an Israeli-based company, explores vision-based technologies for autonomous cars. In fact, Intel and Mobileye unveiled their first autonomous car just three days ago!

Other notable recent acquisitions were made by Baidu (info here) and Ebay (info here).

Such corporate activity is unprecedented for computer vision.

Let’s try to visualise this growth by looking at the following graph (from 2016), showing investments into solely US-based computer vision companies since 2011:

cv-growth-graph
(source)

A clear upward trend can be seen beginning with 2011 when investments were barely above zero. In 2004, when I joined the computer vision club, it would have been even less than that. Amazing isn’t it? Like I said, back then the field very rarely ventured outside of academia. Just compare that with the money being pumped into it now.

Here are some more numbers, this time from venture capital funding:

  • In 2015, global venture capital funding in computer vision reached US$186 million.
  • In 2016, that jumped three-fold to $555 million (source).
  • Last year, according to index.co, investments jumped three-fold again to reach a super cool US$1.7 billion. 

The stand-out from venture capital funding from 2017 was the raising of $460 million from multiple investors for Megvii, a Chinese start-up that develops facial recognition technology (it is behind Face++, a product I hope to write about soon).

Serious, serious money, we’re talking about here.

What the Future Holds for Businesses and Us

Undoubtedly, further growth can be easily predicted for computer vision in the industry. Autonomous cars, commercialisation of drones, emotion detection, face recognition, security and surveillance – these are all areas that will be driving the demand for computer vision solutions for businesses.

Tractica predicts that the market for these solutions will grow to $48.6 billion by 2022. Autonomous cars and robotics will be the major players in these future markets:

cv-forecast-2022
(image taken from here)

What does this mean for us computer vision enthusiasts? Will it be any easier to find those elusive jobs?

In my opinion the state of affairs at our level will not change for a while. A high-level of technical knowledge and understanding backed up by a PhD degree is still going to be the norm for some time to come. Like I mentioned in a previous post, you will need to branch out into other areas of AI to have a decent chance of working on computer vision projects.

Having said that, the situation will slowly start to change once businesses and governments come to realise just some of the things that can be done with the data being acquired by their cameras (more on this in a future post). It’s only a matter of time before this happens, in my opinion, so it’s worth sticking with CV and getting ahead of the crowd now. Investing your time and effort into CV will certainly pay off dividends in the future.

Summary

In this post I presented how much computer vision has grown over the last few years. I looked at some of the recent acquisitions into CV made by big companies such as Apple, Intel and Facebook. I then reviewed the current investments being made into CV and showed that this area is experiencing unprecedented growth. Before 2010, computer vision rarely ventured outside of academia. Now, it is starting to be a viable source of income for businesses around the world. Having said that, the situation for us computer vision enthusiasts will not change for a while. CV jobs will still be elusive. More businesses and governments need to realise that the data being acquired by their cameras can also be mined for information before the effects of this unprecedented growth in CV start to significantly affect us. Thankfully, in my opinion, it’s only a matter of time until this happens so it’s worth sticking with CV and getting ahead of the crowd.

To be informed when new content like this is posted, subscribe to the mailing list:

Deep Learning for Computer Vision with Python Review

Posted on by Zbigatron

In this post I will be reviewing a book called “Deep Learning for Computer Vision with Python“ (DL4CV) that was recently published by Dr Adrian Rosebrock, author of “Practical Python and OpenCV” and most notably the computer vision blog PyImageSearch.

I have already (highly) talked about Dr Rosebrock before on my blog in my post on starting a career in computer vision and I mentioned the fact that PyImageSearch is one of my favourite blogs on the internet. So, it is with great pleasure that I sit down here to write this review.

(Fun fact: I’m writing this review in the wild wilderness of Tasmania. What a beautiful place this is! But, alas, we digress…)

In the first part of this post I will focus on presenting a summary of the book(s) and then in the second part I will give you my thoughts on the work itself. But for those that want the TL;DR version of the review, I’ll give that to you now:

The book is phenomenal. The concepts on deep learning are so well explained that I will be recommending it to anybody not just involved in computer vision but AI in general. If you’re thinking of getting into deep learning for computer vision or wish to fine-tune what you already know, forget about the rest – this is the place to start and finish.

Summary of DL4CV

Due to the huge amount of content that it covers, DL4CV is divided into three volumes: Starter Bundle, Practitioner Bundle, and ImageNet Bundle. Each volume builds on top of the previous one and goes further into the world of deep learning for computer vision. The reason why the volumes are called bundles is that they are each accompanied by additional components such as a downloadable pre-configured Ubuntu virtual machine, source code listings, and access to a companion website. Video tutorials and walkthroughs for each chapter are also advertised to be coming soon.

The Starter Bundle is all about the basics of machine learning, neural networks, convolutional neural networks, and working with datasets. And it truly is a starter bundle because half the book is spent laying down a solid foundation for beginners to deep learning. No knowledge is presupposed (although some experience in computer science or even computer vision I would consider to be advantageous here). Deep learning is also presented on the fundamental level with topics covered such as convolutional neural networks (CNNs), their famous implementations, and the Keras framework. Throughout the book, interesting real-world problems (e.g. breaking captchas) are solved with source code provided and explained at each and every step of the way.

The next volume is the Practitioner Bundle that immerses the reader even further in the world of deep learning. More advanced topics and algorithms are covered such as data augmentation, optimisation methods, and the HDF5 data format. Famous implementations of CNNs are also revisited but in a more in-depth manner. This volume was written for those that want to take computer vision and deep learning (whether it be in academia or the industry) seriously. Once again, practical examples with source code are provided every step of the way.

The final volume is the ImageNet Bundle. The first part of the volume is focused on the ImageNet dataset and the training on it of state-of-the-art CNNs. The second part focuses on even more real-world applications of deep learning and computer vision. Transfer learning and other training techniques are discussed in great detail to the point where readers will be able to reproduce the results seen in seminal deep learning papers and publications. This volume was written for those who want to reach a research level of deep learning in computer vision.

My Thoughts

As I said in the TL;DR section above, this book is phenomenal. And I don’t say things like that lightly – and likewise my words aren’t hot air, either. Prior to returning to the industry this year, my main source of employment was education. I taught and lectured in high schools, primary schools, universities, and privately for eight years. During that time, I acquired a good eye for textbooks that truly give the most to their students in each and every class. A lot of good books exist like this in the fields that I taught in (mathematics, English, philosophy, and computer science). But then sometimes you stumble upon the amazing textbooks – the ones that are just so well-written and structured that they make your job of explaining and helping to assimilate things incredibly easy.

And this is one such book. 

Let me tell you, I know a good educator when I see one – Dr Adrian Rosebrock is one such person. This guy has talent. If I were working at a school or university, I’d hire him without even conducting an interview (well, maybe a quick one over the phone just to make sure he’s not a talented nutcase :P)

I really believe that his talent has produced something unique in the field of deep learning, especially because of the following two characteristics:

  • He understands that when it comes to learning you need to get your hands dirty and do something practical with any newly-acquired theory. That’s the best way to assimilate knowledge. In this respect, all his chapters follow this principle and provide hands-on examples with code to help cement the concepts raised and discussed.
  • His explanations are so ridiculously lucid that he is able to make state-of-the-art academic publications reachable to non-academic people. This is rare. Believe me.

I now work in artificial intelligence in the industry and I am being pushed into a training position in my company. When it’ll come to teaching deep learning, this is the book I will be telling my fellow employees to work through and read with me. And there’s a strong chance that his book will be made into the go-to textbook at universities because of how good it is.

But for that to happen, there is one thing that will need to be touched up. And this thing is my sole criticism of the book.

This criticism is that, in my opinion, there are too many typos (spelling mistakes, missing words, etc.) and grammatical mistakes scattered throughout the book. I understand that such things happen to every writer, but I think that there is an overabundance of them here. I’d say that on average there is one such mistake every few pages. At that rate it can get a little bit frustrating and distracting when you’re trying to focus on the content. When it comes to grammatical mistakes, I’m talking about things like mixing up words such as “affect” and “effect” and “awhile” and “a while”. These creases will need to be ironed out if the book is to be put on shelves in a prominent place in universities and colleges.

However, Dr Rosebrock has provided an easy means to submit mistakes like this to him via the companion website. So, let’s hope that the open-source community will help him out in this respect.

Having mentioned this criticism, I must again underline one thing: this is a unique book and no matter the number of typos and grammatical errors (especially if they will be undoubtedly fixed over time), I hope DL4CV will one day become a classic of deep learning and computer vision. In fact, I’m sure it will.

Summary

In this post I reviewed the book “Deep Learning for Computer Vision with Python” written by Dr Adrian Rosebrock of the PyImageSearch blog. I gave a brief summary of the three volumes and then presented my thoughts on the work as a whole. I mentioned that I think Dr Rosebrock is a talented educator who has written a very good book that explains very difficult concepts exceptionally well. His focus on both theory and implementation is unique and shows that he (perhaps intuitively) understands best-practices in pedagogy. I will be recommending DL4CV to anybody not just involved in computer vision but AI in general. And I hope DL4CV will become a classic textbook at universities.

To purchase “Deep Learning for Computer Vision with Python” or to get more information on it, see the book’s official page.

To be informed when new content like this is posted, subscribe to the mailing list:

cartoon-frustration

Why Image Processing and Computer Vision is so Difficult

Posted on by Zbigatron

This is another post that has been inspired by a question posed in a forum: “What are the open research areas in image processing?”.

My answer? Everything is still an open research area in image processing/computer vision!

But why is this the case? You’d think that after decades of research we’d feel comfortable in saying “this problem here is solved, let’s focus on something else”. In a way we can say this but only for narrow and simple use cases (e.g. locating a red spoon on an empty white plate) but not for computer vision in general (e.g. locating a red spoon in all possible scenarios, like a big box full of colourful toys). I’m going to spend the rest of this post explaining the main reasons behind this.

So, why is computer vision so hard?

Before we dig into what I consider to be the dominant reasons why computer vision is so damn hard, I first need to explain how machines “see” images. When us humans view an image, we perceive objects, people or a landscape. When machines “view” images, all they see are numbers that represent individual pixels.

An example will explain this best. Let’s say that you have a greyscale image. Each pixel, then, is represented by a number usually between 0 and 255 (I’m abstracting here over things like compression, colour spaces, etc.), where 0 is for black (no colour) and 255 is for white (full intensity). Anything between 0 and 255 is a shade of grey, like in the picture below.

grid-intensity-to-numbers
Machines “see” pixels as numbers (pixel boundaries added for clarity).

So, for a machine to garner anything about an image, it has to process these numbers in one way or another. This is exactly what image/video processing and computer vision is all about – dealing with numbers!

Now that we have the necessary background information about computer vision we can move on to the meat of the post: the main reasons behind why computer vision is an immensely hard problem to solve. I’m going to list four such reasons:

  1. Swathes of data
  2. Inherent loss of information
  3. Dealing with noise
  4. Requirements for interpretation

We’ll look at these one at a time.

1. We’re dealing with a heck of a lot of data

As I said above, when it comes to images, all computers see are numbers… lots of numbers! And lots of numbers means a lot of data that needs to be processed to be made sense of.

How much data are we talking about here? Let’s take a look at another example (that once again abstracts over many things such as compression and colour spaces). If you have a greyscale (black & white) image with 1920 x 1080 resolution, this means that your image is described by 2 million numbers (1920 * 1080 = 2,073,600 pixels). Now, if you switch to a colour image, you need three times as many numbers because, typically, when you represent a coloured pixel you specify how much read, blue, and green it is composed of. And then further, if you’re trying to analyse images coming in from a video/camera stream with, say, a 30 frames/sec frame rate (which is a standard frame rate nowadays), you’re suddenly dealing with 180 million numbers per second (3 *2,073,600 * 30 ~= 180 million pixels/sec). That is a lot of data that needs processing! Even with today’s powerful processors and relatively large memory sizes, machines struggle to do anything meaningful with 180 million numbers coming in per second.

2. Loss of Information

Loss of information in the digitising process (going from real life to an image on a machine) is another major player contributing to the difficulty involved in computer vision. The nature of image processing is such that you’re taking information from a 3D world (or 4D if we’re dealing with time in a video stream) and projecting it onto a 2D plane (i.e. a flat image). This means that you’re also losing a lot of information in this process – even though we still have a lot of data to deal with as is, as discussed above.

Now, our brains are fantastic at inferring what that lost data is. Machines are not. Take a look at the image below showing a messy room (not mine, promise!)

cluttered-room

We can easily tell that the large green gym ball is bigger and further away than the black pan on the table. But how is a machine supposed to infer this if the black pan takes up more pixels than the green ball!? Not an easy task.

Of course, you can attempt to simulate the way we see with two eyes by taking two pictures simultaneously and extracting 3D information from these. This is called stereoscopic vision. However, stitching images together is also not a trivial task and is, hence, likewise an open area of research. Further, it too suffers from the other 3 major reasons I discuss in this post.

3. Noise

The digitising process is frequently accompanied by noise. For example, no camera is going to give you a perfect picture of reality, especially when it comes to the cameras located on our phones (even though phone cameras are getting phenomenally good with each new release). Intensity levels, colour saturation, etc. – these will always be just an attempt at capturing our beautiful world.

Other examples of noise are phenomena known as artefacts. These are distortions of images that can be caused by a number of things. E.g. Lens flare – an example of which is shown in the image below. How is a computer supposed to interpret this and work out what is situated behind it? Algorithms have been developed to attempt to remove lens flare from images but, once again, it’s an open area of research.

lens-flare

The biggest source of artefacts undoubtedly comes from compression. Now, compression is necessary as I discussed in this post. Images would otherwise be too large to store, process, and transfer over networks. But if compression levels are too high, image quality decreases. And then you have compression artefacts appearing, as depicted in the image below.

compression-example
The right image has clear compression artefacts visible

Humans can deal with artefacts, even if they dominate a scene, as seen above. But this is not the case for computers. Artefacts don’t exist in reality and are frequently arbitrary. They truly add another level of difficulty that machines have to cope with.

4. Interpretation is needed

Lastly and most importantly is interpretation. This is definitely the hardest thing for a machine to deal with in the context of computer vision (and not only!). When we view an image we analyse it with years and years of accumulated learning and memory (called a priori knowledge). We know, for example, that we can sit on gym balls and that pans are generally used in the kitchen – we have learnt about these things in the past. So, if there’s something that looks like a pan in the sky, chances are it isn’t and we can scrutinise further to work out what the object may be (e.g. a frisbee!). Or if there are people kicking around a green ball, chances are it’s not a gym ball but a small children’s ball.

But machines don’t have this kind of knowledge. They don’t understand our world, the intricacies inherent in it, and the numerous tools, commodities, devices, etc. that we have created over the thousands of years of our existence. Maybe one day machines will be able to ingest Wikipedia and extract contextual information about objects from there but at the moment we are very far from such a scenario. And some will argue that we will never reach a phase where machines will be able to completely understand our reality – because consciousness is something that will always be out of reach for them. But more on that in a future post.

Discussion

I hope I have shown you, at least in a nutshell, why computer vision is such a difficult problem. It is an open area of research and will be for a very, very long time. Ever heard of the Turing test? It’s a test for intelligence devised by the famous computer scientist, Alan Turing in the 1950s. He basically said, that if you’re not able to distinguish between a machine and a human within a specified amount of time by having a natural conversation with both parties, then the machine can be dubbed intelligent.

Well, there is an annual competition called the Loebner Prize that gives away prize money to computer programs deemed most intelligent. The format of the competition is exactly the scenario proposed by Alan Turing: in each round, human judges simultaneously hold textual conversations with a computer program and a human being via a computer. Points are then awarded to how much the machine manages to fool the judges. The top prize awarded each year is about US$3,000. If a machine is able to entirely fool a judge, the prize is $25,000. Nobody has won this award, yet. 

However, there is a prize worth $100,000 that nobody has picked up either. It will be awarded to the first program that judges cannot distinguish from a real human in a Turing test that includes deciphering and understanding text, visual, and auditory input. Once this is achieved, the organisers say that the annual competition will end. See how far away we are from strong intelligence? Nobody has won the $25,000 prize yet, let alone the big one.

I also mentioned above that some simple use cases can be considered solved. I must also mention here that even when use cases appear to be solved, chances are that the speed of the algorithms leave much to be desired. Neural networks are now supposedly performing better than humans in image classification tasks (I hope to write about this in a future post, also). But the state-of-the-art algorithms are barely able to squeeze out ~1 frame/sec on a standard machine. No chance of getting that to work in real-time (remember how I said above that standard frame rates are now at about 30 frames/sec?). These algorithms need to be optimised. So, although the results obtained are excellent, speed is a major issue.

Summary

In this post I discuss why computer vision is so hard and why it is still very much an open area of research. I discussed four major reasons for this:

  1. Images are represented by a heck of a lot of data that machines need to process before extracting information from them;
  2. When dealing with images we are dealing with a 2D reality that has been shrunk from 3D meaning that A LOT of information has been lost;
  3. Devices that present the world to us frequently also deliver noise such as compression artefacts and lens flare;
  4. And the most important hurdle for machines is interpretation: the inability to fully comprehend the world around us and its intricacies that we learn to deal with from the very beginnings of our lives.

I then mentioned the Loebner Prize, which is an AI competition inspired by the Turing test. Nobody has yet won the $25,000, let alone the big one that involves analysing images. I also discussed the need to optimise the current state-of-the-art algorithms in computer vision. A lot of them do a good job but the amount of processing that takes place behind the scenes makes them unusable in real-time scenarios.

Computer vision is definitely still an open area of research.

To be informed when new content like this is posted, subscribe to the mailing list:

sample-gait-recognition

Gait Recognition is Another Form of Biometric Identification

Posted on by Zbigatron

I was watching The Punisher on Netflix last week and there was a scene (no spoilers, promise) in which someone was recognised from CCTV footage by the way they were walking. “Surely, that’s another example of Hollywood BS“, I thought to myself – “there’s no way that’s even remotely possible”. So, I spent the last week researching into this – and to my surprise it turns out that this is not a load of garbage after all! Gait Recognition is another legitimate form of biometric identification/verification. 

In this post I’m going to present to you my past week’s research into gait recognition: what it is, what it typically entails, and what the current state-of-the-art is in this field. Let me just say that what scientists are able to do now in this respect surprised me immensely – I’m sure it’ll surprise you too!

Gait Recognition

In a nutshell, gait recognition aims to identify individuals by the way they walk. It turns out that our walking movements are quite unique, a little like our fingerprints and irises. Who knew, right!? Hence, there has been a lot of research in this field in the past two decades.

There are significant advantages of this form of identity verification. These include the fact that it can be performed from a distance (e.g. using CCTV footage), it is non-invasive (i.e. the person may not even know that he is being analysed), and it does not necessarily require high-resolution images for it to obtain good results.

The Framework for Automatic Gait Recognition

Trawling through the literature on the subject, I found that scientists have used various ways to capture people’s movements for analysis, e.g. using 3D depth sensors or even using pressure sensors on the floor. I want to focus on the use case shown in The Punisher where recognition was performed from a single, stationary security camera. I want to do this simply because CCTV footage is so ubiquitous today and because pure and neat Computer Vision techniques can be used on such footage.

In this context, gait recognition algorithms are typically composed of three steps:

  1. Pre-processing to extract silhouettes
  2. Feature extraction
  3. Classification

Let’s take a look at these steps individually.

1. Silhouette extraction

Silhouette extraction of subjects is generally performed by subtracting the background image from each frame. Once the background is subtracted, you’re left with foreground objects. The pixels associated with these objects can be coloured white and then extracted.

Background subtraction is a heavily studied field and is by no means a solved problem in Computer Vision. OpenCV provides a few interesting implementations of background subtraction. For example, a background can be learned over time (i.e. you don’t have to manually provide it). Some implementations also allow for things like illumination changes (especially useful for outdoor scenes) and some can also deal with shadows. Which technique is used to subtract the background from frames is irrelevant as long as reasonable accuracy is obtained.

silhouette-extraction
Example of silhouette extraction

2. Feature extraction

Various features can be extracted once we have the silhouettes of our subjects. Typically, a single gait period (a gait cycle) is first detected, which is the sequence of video showing you take one step with each of your feet. This is useful to do because your gait pattern repeats itself, so there’s no need to analyse anything more than one cycle.

Features from this gait cycle are then extracted. In this respect, algorithms can be divided into two groups: model-based and model-free.

Model-based methods of gait recognition take your gait period and attempt to build a model of your movements. These models, for example, can be constructed by representing the person as a stick-figure skeleton with joints or as being composed of cylinders. Then, numerous parameters are calculated to describe the model. For example, the method proposed in this publication from 2001 calculates distance between the head and feet, the head and pelvis, the feet and pelvis, and the step length of a subject to describe a simple model. Another model is depicted in the image below:

Example-of-gait-model
An example of a biped model with 5 different parameters as proposed in this solution from 2012

Model-free methods work on extracted features directly. Here, undoubtedly the most interesting and most widely used feature extracted from silhouettes is that of the Gait Energy Image (GEI). It was first proposed in 2006 in a paper entitled “Individual Recognition Using Gait Energy Image” (IEEE transactions on pattern analysis and machine intelligence 28, no. 2 (2006): 316-322).

Note: the Pattern Analysis and Machine Intelligence (PAMI) journal is one of the best in the world in the field. Publishing there is a feat worthy of praise. 

The GEI is used in almost all of the top gait recognition algorithms because it is (perhaps surprisingly) intuitive, not too prone to noise, and simple to grasp and implement. To calculate it, frames from one gait cycle are superimposed on top of each other to give an “average” image of your gait. This calculation is depicted in the image below where the GEI for two people is shown in the last column.

The GEI can be regarded as a unique signature of your gait. And although it was first proposed way back in 2006, it is still widely used in state-of-the-art solutions today.

gait-energy-image
Examples of two calculated GEIs for two different people shown in the far right column. (image taken from the original publication)

3. Classification

Once step 2 is complete, identification of subjects can take place. Standard classification techniques can be used here, such as k-nearest neighbour (KNN) and the support vector machine (SVM). These are common techniques that are used when one is dealing with features. They are not constrained to the use case of computer vision. Indeed, any other field that uses features to describe their data will also utilise these techniques to classify/identify their data. Hence, I will not dwell on this step any longer. I will, however, will refer you to a state-of-the-art review of gait recognition from 2010 that lists some more of these common classification techniques.

So, how good is gait recognition then?

We’ve briefly taken a look at how gait recognition algorithms work. Let’s now take a peek at how good they are at recognising people.

We’ll first turn to some recent news. Only 2 months ago (October, 2017) Chinese researchers announced that they have developed the best gait recognition algorithm to date. They claim that their system works with the subject being up to 50 metres away and that detection times have been reduced to just 200 milliseconds. If you read the article, you will notice that no data/results are presented so we can’t really investigate their claims. We have to turn to academia for hard evidence of what we’re seeking.

“Gaitgan: invariant gait feature extraction using generative adversarial networks” (Yu et al., IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp. 30-37. 2017) is the latest top publication on this topic. I won’t go through their proposed algorithm (it is model-based and uses the GEI), I will just present their results – which are in fact quite impressive.

To test their algorithm, the authors used the CASIA-B dataset. This is one of the largest publicly available datasets for gait recognition. It contains video footage of 124 subjects walking across a room captured at various angles ranging from front on, side view, and top down. Not only this, but walking is repeated by the same people while wearing a coat and then while wearing a backpack, which adds additional elements of difficulty to gait recognition. And the low resolution of the videos (320×240 – a decent resolution in 2005 when the dataset was released) makes them ideal to test gait recognition algorithms on considering how CCTV footage has generally low quality also.

Three example screenshots from the dataset is shown below. The frames are of the same person with a side-on view. The second and third image shows the subject wearing a coat and a bag, respectively.

casia-example
Example screenshots from the CASIA B dataset of the same person walking.

Recognition rates with front-on views with no bag or coat linger around 20%-40% (depending on the height of the camera). Rates then gradually increase as the angle nears the side-on view (that gives a clear silhouette). At the side-on view with no bag or coat, recognition rates reach an astounding 98.75%! Impressive and surprising.

When it comes to analysing the clips with the people carrying a bag and wearing a coat, results are summarised in one small table that shows only a few indicative averages. Here, recognition rates obviously drop but the top rates (obtained with side-on views) persist at around the 60% mark.

What can be deduced from these results is that if the camera distance and angle and other parameters are ideal (e.g. the subject is not wearing/carrying anything concealing), gait recognition works amazingly well for a reasonably sized subset of people. But once ideal conditions start to change, accuracy gradually decreases to (probably) inadequate levels.

And I will also mention (perhaps something you may have already garnered) that these algorithms also only work if the subject is acting normally. That is, the algorithms work if the subject is not changing the way he usually walks, for example by walking faster (maybe as a result of stress) or by consciously trying to forestall gait recognition algorithms (like we saw in The Punisher!).

However, an accuracy rate of 98.75% with side-on views shows great potential for this form of identification and because of this, I am certain that more and more research will be devoted to this field. In this respect, I will keep you posted if I find anything new and interesting on this topic in the future!

Summary

Gait recognition is another form of biometric identification – a little like iris scanning and fingerprints. Interesting computer vision techniques are utilised on single-camera footage to obtain sometimes 99% recognition results. These results depend on such things as camera angles and whether subjects are wearing concealing clothes or not. But much like other recognition techniques (e.g. face recognition), this is undoubtedly a field that will be further researched and improved in the future. Watch this space.

To be informed when new content like this is posted, subscribe to the mailing list:

infrared-image-of-human

Lie Detection with Thermal Imaging – A Task for Computer Vision

Posted on by Zbigatron

Can thermal imaging detect if you’re lying or not? It sure can! And are there frightening prospects with respect to this technology? Yes, there are! Read on to find out what scientists have recently done in this area – and all using image processing techniques.

Thermal Imaging

Thermal imaging (aka infrared thermography, thermographic imaging, and infrared imaging) is the science of analysing images captured from thermal (infrared) cameras. The images returned by these cameras capture infrared radiation not visible to the naked eye that are emitted by objects. All objects above absolute zero (-273.15 °C or −459.67°F) emit such radiation. And the general rule is that the hotter an object is, the more infrared radiation it emits.

There has been some amazing work done recently by scientists with respect to thermal imaging and deception detection. These scientists have managed to construct sophisticated lie detectors with their thermal cameras. And what is interesting for us is that these lie detectors work by using many standard Computer Vision techniques. In fact, thermal imaging is a beautiful example of where Computer Vision techniques can be used on images that do not come from “traditional” cameras.

This post is going to analyse how these lie detectors work and underline the Computer Vision techniques that are being used by them. The latter part of the post will extrapolate how thermal imaging might impact us in the future – and these predictions are quite frightening/exciting, to say the least.

Lie Detectors

The idea behind lie detectors is to detect minor physiological changes, such as an increase in blood pressure, pulse or respiration, that can occur when we experience a certain anxiety, shame or nervousness when dropping a fib (aka telling a porky, taking artistic license, being Tony Blair, etc.). These are such slight physiological changes in us that instruments that measure them need to be very precise.

We’ve all seen polygraphs being used in films to detect deception. An expert sits behind a machine and looks at readings from sensors that are connected to the person being interrogated. Although accuracy is said to be at around 90% in detecting lies (according to a few academic papers I studied), the problem is that highly trained experts are required to interpret results – and this interpretation can take hours in post-interview analysis. Moreover, polygraph tests require participants’ cooperation in that they need to be physically connected to these sensors. They’re what’s called ‘invasive’ procedures.

lie-detecting-Keeler
A polygraph test being conducted in 1935 (Image source: Wikipedia)

Thermal imagery attempts to alleviate these problems. The idea behind them is to detect changes in the surface temperature of the skin caused by the effects of lying. Since all one needs is a thermal camera to observe the participant, they’re non-invasive procedures. And because of this, the person being interrogated can be oblivious to the fact that he’s being scrutinised for lying. Moreover, the process can be automated with image/video processing algorithms – no experts required for analysis!

Computer Vision in Thermal Imaging

Note: Although I gloss over a lot of technical details in this section, I still assume here that you have a little bit of computer vision knowledge. If you’re here for the discussion on how thermal imagery could be used in the future, skip to the next section. 

There are some interesting journal papers on deception detection using thermal imagery and computer vision algorithms. For example, “Thermal Facial Analysis for Deception Detection” (Rajoub, Bashar A., and Reyer Zwiggelaar. IEEE transactions on information forensics and security 9, no. 6 (2014): 1015-1023) reports results of 87% on 492 responses (249 lies and 243 truths). Techniques such as machine learning were used to build a model of deceptive/non-deceptive responses. The models were then utilised to classify responses.

But I want to look at a paper published internally by the Faculty of Engineering at the Pedagogical and Technological University of Colombia in South America. It’s not a very “sophisticated” publication (e.g. axes are not labelled, statistical significance of results is not presented, the face detection algorithm is a little dubious, etc.) but the computer vision techniques used are much more interesting to analyse.

The paper in question is entitled “Detection of lies by facial thermal imagery analysis” published this year (2017) by Bedoya-Echeverry et al. The authors used a fairly low-resolution thermal camera (320×240 pixels) to obtain comparable results to polygraph tests: 75% success rate in detecting lies and 100% success rate in detecting truths.

I am going to work through a simplified version of the algorithm presented by the authors. The full version involves a few extra calculations but what I present here is the general framework of what was done. I’ll give you enough to show you that implementing a thermal-based lie detector is a trivial task (once you can afford to purchase the $3,000 thermal camera).

This simplified algorithm can be divided into two stages:

  1. Face detection and segmentation
  2. Periorbital area detection and tracking

Let’s work through these two stages one-by-one and see what a fully-automated lie detector based on Computer Vision techniques can look like.

1. Face detection and segmentation

For face detection the authors first used Otsu’s method on their black and white thermal images. Otsu’s method takes a greylevel (intensity) image and reduces it to a binary image, i.e. an image containing only two colours: white and black. A pixel is coloured either white or black depending on whether it falls below or above a certain dynamically calculated threshold. The threshold is chosen such that it minimises the intra-class variance of the intensity values. See this page for a clear explanation on how this is done exactly.

When the binary image has been produced, the face is detected by calculating the largest connected region in the image.

face-detection-from-IR-image
(Image adapted from original publication)

Note: functions for Otsu’s method and finding the largest connected components are all available in OpenCV and are easy to use.

2. Periorbital area detection and tracking

Once the face has been detected and segmented out, the next step is to locate the periorbital region, which is the horizontal, rectangular region containing your eyes and top of your nose. This region has a high concentration of veins and arteries and is therefore ideal for scrutinising for micro-temperature changes.

The periorbital region can be found by dividing the face (detected in step 1) into 4 equally-spaced horizontal strips and then selecting the second region from the top. To save having to perform steps one and two for each frame, the KLT algorithm is used to track the periorbital area between frames. See this OpenCV tutorial page for a decent explanation of how this tracking algorithm works. It’s a little maths intensive – sorry! But you can at least see that it’s also easy to implement.

Temperature readings are then made from the detected region and an average calculated per frame. When the average temperature (i.e. pixel intensity) peaks during the answering of a question, the algorithm can deduce that the person is lying through their teeth!

That’s not that complicated, right? Even though I simplified the algorithm a little (a few additional calculations are performed to assist tracking in step 2), the gist of it is there! Here you have a fully-automated, non-invasive lie detector that uses Computer Vision techniques to get results comparable to polygraph tests.

What the Future Holds in Lie Detection and Thermal Imagery

Now, let’s have a think about what a non-invasive lie detector could potentially achieve in the future.

Lie detectors are all about noticing micro-changes in blood flow, right? Can you imagine such a system tracking you around the store to gauge which products get you a little bit excited? Results can be instantly forwarded to a seller who now has the upper hand over you. Nobody will be safe from second-hand car dealers any more. To confirm anything, all he has to do is ask if you like a certain car or not.

What about business meetings? You can put in a cheeky thermal camera in the corner of a meeting room and get live reports on how clients are really feeling about your pitch. Haggling will be a lot easier to deal with in this way if you know what your opponent is truly thinking.

And what about poker? You will be able to beat (unethically?) that one friend who always cleans up at your “friendly” weekend poker nights.

The potential is endless, really. And who knows?! Maybe we’ll have thermal cameras in our phones one day, too? Computer vision will definitely be a powerful tool in the future 🙂

What other uses of deception detection using thermography can you think of?

Summary

Traditionally, lie detection has been performed using a polygraph test. This test, however, is invasive and needs an expert to painstakingly analyse results from the various sensors that are used. Digital thermography is looking like a viable alternative. Scientists have shown that using standard computer vision techniques, deception detection can be non-invasive, automated, and get results comparable to polygraph tests. Non-invasive lie detectors are a scary prospect considering that they could track our every move and analyse all our emotions in real time.

To be informed when new content like this is posted, subscribe to the mailing list:

do-not-give-up

How Can I Start a Career in Computer Vision?

Posted on by Zbigatron

This post has been inspired by a question someone asked in a forum. This person was a new but competent programmer who was trying to move into Computer Vision (CV) in the industry. However, he rightly noticed that “most of the job requirements [in computer vision] are asking for a PhD”.

Indeed, this is true. And because of this, finding a job in computer vision in the industry is very difficult. Mainly because computer vision (e.g. video analytics) hasn’t caught on yet as a meaningful source of data or income for companies. I think it will sooner or later (more on this in a future post) but the reality is that such jobs are rare and hence why companies can afford to advertise and be picky for people with a PhD.

So, if you’re in a situation where you can’t land a job in CV and especially if you’re without a PhD, don’t give up. Computer vision is a fascinating field to work in and it will only get bigger with time. It’s worth fighting on.

Here are three things you can do to improve your chances of landing that job that you really want.

Read Books, Tutorials, Publications, and Blogs

This is obvious but needs to be said. Keep reading up on the field, keep fine-tuning your skills and knowledge in CV. You need to show your potential employer that you know the field of CV exceptionally well. Read important books that are or have been published. Some you can find in your library, some come in PDF format. For example, Neural Networks and Deep Learning is a great book on the hot topic of neural networks that is available online free of charge.

Work through the tutorials available on the OpenCV page. There’s plenty there on machine learning, photo processing, object detection, etc. to keep you busy for months! The idea is to get so good at CV to be able to instantly see a solution to an image/video processing problem. You need to shine at those job interviews.

Follow blogs on Computer Vision. Two of my favourite are PyImageSearch and Learn OpenCV. These guys are regularly posting stuff that will fascinate anybody with a passion for Computer Vision. In fact, PyImageSearch is so well-written that it put me off from starting this blog for a while.

Consider also looking into academic publications. These can be daunting, especially if you don’t have a background in research. But focus initially on the seminal papers (more on this in a future post) and try to get the gist of what the scientists are saying. You can usually pick up small bits and pieces here and there and implement simplified versions of them.

Side Projects

One thing that did get me a lot of attention were my side projects. Side projects show people where your passions lie. And passion is something that a lot of companies are looking for. Believe me, if you came to my company and I was asked to interview you to join the Computer Vision team, your side projects would be one of the first things I’d be looking at.

So, get stuck into a few of these to show that you love the area and you do this kind of stuff for fun. Get a Raspberry PI going with a camera and build your own security system via motion detection, for example. Or get a drone for your Raspberry Pi and camera and be creative with it. Then list these side projects at the end of your CV. If you’re truly passionate about Computer Vision, you will get noticed sooner or later.

Branch out into other areas of Artificial Intelligence

What I decided to do when my job hunting wasn’t going too well was to aim for jobs in other areas of AI rather than just Computer Vision. It involved me having to pick up additional knowledge in fields I wasn’t too familiar with (e.g. Robotic Process Automation) but the amount of jobs in these areas is much larger. This tactic proved successful for me. I ended up picking a company (yes, I was spoilt for choice in the end!) that had interesting clients and it was only a matter of time before opportunities for computer vision projects came along that we were all pushing for.

An Inspirational Story

If you are feeling down about your job searching or if you’re wondering whether CV is a viable place to aim for in the job market, here is a truly inspirational story out of India (from PyImageSearch – I told you it was a well-written blog!). It’s about a fellow who really wanted to work in CV but was at a disadvantage because he came from a low-income family. But he didn’t give up and put the hard work in. Today he is working on AI solutions for drones for a company in India. Moreover, with his salary he can support his family, has paid off all his debts, and is working in a field he absolutely loves!

Summary

Finding a job in Computer Vision is difficult. Most companies are advertising for people with a PhD. But there are things you can do to boost your chances of landing that job you really want. For example, you can keep your CV skills sharp by continually reading up on the subject. You can also work on side projects in CV and you can try branching out into other areas of AI to broaden the scope of projects you are qualified for. Don’t give up on your quest because CV is a field that is going to grow – it’s only a matter of time before the industry catches on to the amazing things CV can do with their video data.

To be informed when new content like this is posted, subscribe to the mailing list: