We really needed an excuse to upgrade all of our compute but we first had to prove to ourselves that it was worth it. Many you may be uploading videos to your preferred AI platform and running amazing analytics. Even more of you are running out of tokens or too much lag or even the dreaded file size too large problem. We have visited each of those pain islands. The next maturation is that you are downloading individual ai programs to your iPhone that turn the camera of the iPhone inward and provide decent analysis. We did that and had several mini applications on our phones doing a myriad of things. We even created a custom iPhone app in SWIFT to analyze golf swing videos. We were not alone - today there are 50 some iPhone apps for measuring and visualizing golf swing dynamics. Some of those have now turned into paid apps and more will follow.
In the pipeline of video analysis economics is a constant concern. We have now turned the economic lens from the cloud to the resting compute cycles on our little MacBook Air and are taking advantage of the long used command ‘caffeinate’ to keep the training models running while we sleep (does the MacBook ever sleep is the question).
Here are several inference runs on our custom model designed for Edge Flow Analytics. The impetus of the custom model is that we are deriving a different set of analysis that are not your standard boxscore fare. We are looking at ball flow as it leads to scoring and are using 2 years of training data to hypothesize where defence should be not who is guarding who. Shifting the defensive model from on player dynamics to on space (space on the floor dynamics).
The most interesting discovery so far is that since we are using two years of data we were able to dim many of the bright spots on the floor and use a new formative Bayesian technique to smooth out areas of the floor that the camera did not have a great angle of based on previous camera angle movements, or in the case of Championship games we were able to stitch other camera angles onto game play footage. We are most excited by the gamification of this all. We are able to produce clips derived from thousands of positions and create a video of bayesian ball placement from offensive positions so that we can teach better defensive positioning and ultra realistic clips that can be snipped to 15 seconds. In other words 2 years of data down to what looks like a real game - but the game did not happen - it is predictive of what will/could happen based on thousands of examples.
We will provide some of the underlying training performance metrics in the next post.