Video wird geladen...
Video konnte nicht geladen werden
Cool video capturing how computer vision can accurately analyze a random basketball game. Some stuff to notice: 1) This is a random court, not one it was trained on. 2) The camera is moving, meaning the model has to calibrate where all the lines are accordingly. 3) It detects... show more
14,619 Aufrufe • vor 14 Tagen •via X (Twitter)
16 Kommentare

@SportsVisioAI Ex-college player here but beginning to coach my son. I've dreamed of this. Dual monitors, one with the video the other top down 2d view of plays executing all live with stats being recorded. Excited!

@SportsVisioAI Follow us and it will be ready to go when your son is! The AI just keeps getting faster and cheaper as we build it, hoping to be realtime by sometime next year Lord willing.

@SportsVisioAI This is good stuff.

@SportsVisioAI Is this open source?

@SportsVisioAI No, but we have contributed some code to open source. We spent millions to get this far, building a business around it. But the goal is to be much more affordable than paying humans, only a dollar and change per player or even less!

@SportsVisioAI how many frames per second to get this result?

@SportsVisioAI We like 30 fps but can go down to 24 without much degradation. Below that it gets dicey and you start to really see the accuracy fall off as you head downwards. 1080p video is also ideal but we can handle 720, again below that you see real performance impact

@SportsVisioAI thx. sorry i should have asked how many frames are you actually achieving on analysis/inference? assume with 10 players and a ball you are not at real time?

@SportsVisioAI Sorry I don’t understand the question. We are processing ALL the frames. An hour game takes us about 6 hours to process (down from 22!) and we are trying to get to realtime next year.

@SportsVisioAI Thank you that’s what I was trying to ask. Appreciate it

@SportsVisioAI how do you keep the track the metrics of each player and are you storing them in a relational database?

@SportsVisioAI How do you do ball possession to players? Is it box overlap or distance to nearest player center? How would you tackle adding the ball to the homography chart?

We built a ton of custom tracking algorithms. Player tracking, ball tracking, possession tracking, jersey detection, action detection, line detection, etc. Each one was specially trained for that task against our training corpus. We have over 30,000 full games with video and over 100,000 in total now just for basketball. We actually have a crude model showing it on the homography chart, the player with the yellow circle around them.

Love it that people are seeing and brining full potential of ML and CV, Building Similar for Canadian practice games/ sessions, Just reaching good level of Detection + Segmentation + OCR + Team Association actually took me around 3-4 months as a Solo Researcher and Developer. Built it by finetuning the fully open sourced model within our domain.. Also for saving compute resources, I use sam segmentation to segment the Court for scoring system just before the x second of identifying a Make / Miss event.. : ) .. Rn not keeping track of positions.. But I just explored the PTZ for panning camera, and Yes we can create a whole court Panorma and use SIFT for player Position/ feet detection. Its gonna be fun

@SportsVisioAI Got anything for field hockey? I’m two years out from needing this.

@SportsVisioAI Not yet but always interested in what new sports to pursue. We do volleyball also currently, and have prototypes for baseball and tennis. Have a lot of requests for lacrosse and flag football, and of course soccer comes up a lot too
