When you introduce advanced visual technology within an already existing lucrative market, you have the potential to create an explosive reaction. That is exactly the model SportLogiq is using, introducing image analysis into the multi-million market of sports analytics. Estimated at $125 million in 2014, sports analytics is anticipated to reach $4.7 billion by 2021. Mostly done manually using banks of humans processing hours of videos, it is now on the verge of being completely disrupted, thanks to the efforts of the Marc Cuban-backed, Montreal-based startup.

We caught up with CTO and co-founder Mehrsan Javan to learn more:

– A little about you, what is your background?

I am currently the CTO and a co-founder of SportLogiq, a computer vision based sports analytics startup based in Montreal. My passion for image analysis and

Mehrsan Javan Co-Founder and CTO at SPORTLOGiQ
Mehrsan Javan Co-Founder and CTO at SPORTLOGiQ

computer vision comes from my undergrad study when I became interested in biomedical image analysis. I worked a couple of years as an automation and control system engineer while working toward my master degree in electrical engineering. I did my PhD. in McGill University focusing on computer vision and content-based video retrieval and have been involved in many research projects about vision based tracking and human activity recognition in videos. This is where the core technology behind SportLogiq was born. I am still involved with academia and affiliated with McGill University as an adjunct professor.

After trying to commercialize some technologies developed in my PhD. study, I eventually co-founded SportLogiq with my brilliant co-founder Craig Buntin, a former Olympic athlete.
– Explain SportLogIQ. What does it solve and how does it work?
SportLogiq provides computer vision solutions for comprehensive sports statistics and offers an advanced sports analytics platform. This includes tracking sports players, describing their activities, and analyzing the game strategies, players’ performance, and advanced statistics such as scoring chance. These solutions produce accurate and high-level statistics in sports analytics which are extremely useful for both teams and media. What SportLogiq does actually changes the way that the games are described, analyzed, and the way players are being evaluated and scouted.
In a most abstract and simple description, there are two processing engines which form the basis of the SportLogiq sports analytics platform. The first one, the computer vision engine that is basically a data acquisition tool which extracts all information from video feeds. The computer vision engine uses the feed from a single camera to track all players and describes their activities in a game, which are essentially about who is where and doing what at any given time during a game. This information is then used by the game analytics engine which builds a statistical game model in order to generate the high-level descriptive and predictive statistics of a game. This is how SportLogiq works, a pretty simple and straightforward procedure.
Computer vision is always being considered as a trivial thing by non-expert people because the visual interpretation in human brain is being done unconsciously and hence, no one notices how complicated that process is – this is always an issue when we want to sell computer vision technology to the public, one part of my job is to teach others and set their expectations about high-level computer vision stuff.
– Can it do this real-time? Or is there a delay?
First of all, real-time is a vague term because it has to be defined in the context. For example, a real-time statistics for a team might be right after each period, for broadcasters might be within couple of minutes latency in a game, and for digital media and post-game analytics might be couple of hours after a game.  All of the computer vision processing are intense and hence, there is always a latency between the currently observed video frame and the last processed one. At the current stage, we can provide statistics and advanced analytics within 5 minutes delay. This processing time is getting improved and we foresee real-time analytics (maybe with a couple of milliseconds latency) in a very near future.
The system uses self-calibrating tool to understand what part of the field it is currently seeing.
The system uses self-calibrating tool to understand what part of the field it is currently seeing.
– Can it use regular broadcast tv or does it need a special feed?
The competitive advantage of SportLogiq is that we can process any game recorded using any standard camera pull stats out of it. In other words, there is no need to install specific hardware or to use a specific camera. We can process the broadcast feed, exactly the same one which a viewer watches on TV. This is the whole idea behind SportLogiq to generate advanced statistics from currently available hardware. Our technology will revolutionize everything in sport analytics, even more than what multiple-fixed camera tracking systems did almost ten years ago.
– Hockey first and then other sports? Which ones do you have in mind?
We started with hockey because of hundred different reasons, among them is that we are Canadian and we love hockey and we understand it well! Hockey is among four biggest sports in North America and there is a big opportunity in this market. I am not talking only about the National Hockey League (NHL), we can expand the technology to other hockey leagues and even college games. Once the technology is established in hockey, we will extend it to other sports, the next immediate one would be another indoor sport. I’d rather prefer not to talk directly about the next sport, but I think it is not hard to guess which one would be the next… we intend to broaden our scope to include basketball, soccer, and other sports.
– Will it consider natural elements ( like rain, wind, heat) in its results ?
This is a very smart question. Changes in the environment have effects on both computer vision and game analytics part.  Computer vision problems are still among the most complicated problems in the computer science. Despite the huge amount of work which has been done in the past, a universal solution that works in the wild is still a big challenge in computer vision. This is still and ongoing research in this field and as of today, there are good solutions for those problems. On the other hand, sports are almost restricted environments and it is somehow easier to develop robust solutions for those situations. Our computer vision solutions are robust enough for outdoor environments, but it is good to remind that hockey is an indoor sport.
From the game analytics and player performance evaluation perspective, the answer is different. IT is true that natural elements may affect players’ performance in a game, but the question is whether to include those effects in our analysis or not. This is a traditional question in machine learning, what kind of information shall be considered in order to have a good understanding about a phenomenon. There are millions of things that can affect a player performance in a game, and including everything in the analysis is the thing which is referred to as the curse of dimensionality.
Since we are working in an indoor environment, there are not many natural elements and hence, we have no evidence about their effects on the way that a game is being played. We will see those things once we move to soccer analysis…
– Can the technology be expanded to usage outside of sports? For example, crowd analysis and patterns or animal behaviors?
Defiantly yes, sport is an application of our computer vision based human activity understanding. Most of our computer vision algorithms are not completely application specific and can be applied to other computer vision problems in different domains. As an example, the parts of our original technology which comes from my PhD thesis, were originally developed for human activity recognition and abnormal behavior detection in security and surveillance videos; an application which is completely different from sport analytics and we adopted the same technology for our current application.
SportLogiq identifies and tracks each individual player throughout the game.
SportLogiq identifies and tracks each individual player throughout the game.
– Will it be able to predict accurately which teams will win based on past patterns?
The short answer is yes, we can provide predictive statistics and estimate the scoring chance. The long answer is prediction is sophisticated and still is one of the most challenging problems in artificial intelligence. This becomes more complicated and less accurate when it involves human behavior. As a simple example, imagine a player gets injured early in the season. Then none of the previous predictions are valid anymore and it is required to predict things by excluding some independent variables.  Given the amount of data that we have, we can predict many things with some degrees of confidence. In overall, prediction based on team level data is more confident compared to performance prediction for individual players.
– Also, can it detect cheating ( like drugs)  by detecting unusual patterns?
To be honest, I am not sure about it and this is something that needs to be investigated. Obviously we can identify unusual/rare patterns, but we cannot explain what causes those abnormalities. An explanation for an abnormality is far beyond a computer’s capability.
– Are you finding that you can track stats that were impossible for a human to do previously and learn more about a players abilities?
There is a significant difference between human and a machine. We are very good at qualitative measurements and high-level inference while we are incapable of doing quantitative measurements. A very simple example is that a human can say someone is a fast skater, but cannot say how fast the player skates or who is faster than who, without a side by side comparison. In SportLogiq, everything is quantitatively measured and then analyzed and described to be informative and understandable for a human. Or another example is the “exact” location of a player on the ice, which is impossible for a human to do. In each hockey game we roughly measure 2500 different game events with their precise locations in real world coordinates systems, those can never be measured and analyzed by a human. It gives us the possibility to compare players from across the National Hockey League, and provide quantitative insights to the human to make better decisions or have better understandings about the game.
– Who is your typical customer?
As of now our clients are broadcast media and professional teams. We have an amazing business team and they are trying to dominate both markets.
SportLogiq tracking technology at work
SportLogiq tracking technology at work
– Tell us about your consumer offering? Will I be able to use SportLogiq to track my kids progress?
This is the big objective that we have, introducing a sports analytics platform to be used not only by the professional leagues  but also every amateur sports player all around the world. Record a video through our app on your mobile device and see the stats in real-time, and parents can watch their kid’s progress. This is what we will work on to offer to consumers in near future. This will also completely change the way that players are being scouted. What we are doing will potentially throw thousands of players into a giant system where scouts can automatically say: well these players are developing well, these players have talent, and we should keep an eye on these kids. This will affect the future of millions of amateur athletes all around the world and give everyone an equal opportunity to get recognized.
– What would you like to see SportLogiq offer that technology cannot yet deliver?
Despite all the hype, the technology is much further off than what people think and there are still many unsolved challenges. Computer vision and artificial intelligence are still new technology fields and they have a long way ahead to become mature. What makes me excited is to see that we can understand and describe visual contents, do high-level inferences, and understand the context – similar to what human does and of course without using a huge labeled training data in the supervised learning paradigm. As a side note, supervised learning algorithms are only a small subset of learning methods and despite lots of progress in this field, we are far behind what human learns from patterns, specifically in an unsupervised fashion. Once we get to that point it will change everything for humans and makes our life much easier and more comfortable, and somehow uncomfortable…


Author: Paul Melcher

Paul Melcher is the founder of Kaptur and Managing Director of Melcher System, a consultancy for visual technology firms. He is an entrepreneur, advisor, and consultant with a rich background in visual tech, content licensing, business strategy, and technology with more than 20 years experience in developing world-renowned photo-based companies with already two successful exits.

1 Comment

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.