AI Punch Count : Canelo Alvarez vs Gennady Golovkin, II

*DeepStrike Video Counter* This is the first version our punch counter visualization. It’s still work in process, we’d love to get some feedback on how we can improve it. Note that actually following and counting punches in real-time for most humans is extremely difficult due to the fast nature of the sport. So to check the classifications we suggest pausing the video and pressing the , and . or the angle bracket buttons. That way you can step through the video in YouTube frame-by-frame to better see the classifications extracted. From a single-angle camera feed, the accuracy will be effected, however in a statistical sense it is still very good. See notes on Accuracy and Impact Scale below. We’re a small team of engineers and boxing fans working on this, and DeepStrike is very much still in Beta mode, so in the video you’ll see occasional jittering of the AI and tracker. Any feedback hints suggestions input etc will be much appreciated! :D 0:03 - Round 1 3:17 - Round 2 6:45 - Round 3 10:11 - Round 4 13:39 - Round 5 17:04 - Round 6 20:31 - Round 7 23:55 - Round 8 27:22 - Round 9 30:47 - Round 10 34:11 - Round 11 37:37 - Round 12 40:47 - Full Fight Stats 41:09 - Winner Declaration #canelo #ggg #canelovsggg #saulalvarez #gennadiygolovkin #boxing #deepstrike #boxeo *Impact Scale:* Not all landed punches are equal. Some punches may land perfectly with devastating visible effect, and others do make contact yet are partially defended or imperfect in some aspects and therefore have limited impact on the target. We found early on that simply counting punches alone, can in some fights be misleading in terms of being descriptive of the fight. For a meaningful metric, one must take into account the quality with which punches are landing, splitting the counts into Jabs and Non-Jabs is not sufficient and can occasionally be very misleading. The impact scale in brief goes through Min, Low, Mid, High, Max. The Max and High Impact categories are for punches that lands cleanly, accurately, undefended, with power, and has clear visible effect on target. Mid and Low Impact are for punches that landed but are imperfect in one or more ways such as being partially defended with blocks, pulls, rolls, or simply making contact without much power. The Min Impact category is for punches that connect, but are imperfect to the extent that it is debatable whether or not they should even be included as a landed punch. There are in general wildly different ideas and thresholds in the boxing community for when a punch should be considered landed, and so we therefore suggest paying more attention to the impact counters instead of the overall count that sums everything from Min to Max. The AI is being trained on data from hundreds of thousands of samples tagged by boxing data annotators, it learns to mimic their instructions, and as such will be subject to any potential bias within the data-set. In addition there is a level of subjectivity to the scale, and even highly skilled boxing annotators reviewing video footage in slow-motion don’t always tag a punch into the same category. As such the impact category to which a punch is associated shouldn’t be considered a ground truth, rather think of it as a stochastic variable that gives a strong indication of the distribution of the quality and impact of the punches that each fighter is landing. *Accuracy:* It’s important to note that DeepStrike can only count what is visible in the frames. When counting on a single dirty feed video, punches for which the target is occluded will be marked as Missed even if they may actually have Landed in some cases. This tends to be more of an issue for body-shots where occlusion happens more frequently than head-shots. DeepStrike can run directly on multiple iso feeds from several cameras and merge the output for perfect visibility so to avoid any issues of occlusion, though this is mainly for broadcasters as typically the iso feeds from each camera is not generally available. As a concept, DeepStrike counts on a maximum confidence basis. So if something is unclear or can’t be seen, it won’t be counted. This also means that counts will be lower in video material of lower quality as the outcome of some punches may then not be visible due to motion blur or limited spatio-temporal resolution in the source video feed. For Throwback Thursday stats published we run them all with human supervision to check and correct any obvious errors. Since most of the fights requested are very controversial, we feel it’s important that the stats are of high quality. The occlusion limitation with single-feed video is however the same for both human and machine, and so it’s not something the human checks can “fix“.