Real-Time Ball Possession Tracking
Professional Sports Analytics Provider
Automated ball possession detection in live basketball broadcasts, replacing manual video review.
Tracking which player holds the basketball across thousands of broadcast frames per game required a team of analysts reviewing footage manually. Accuracy hovered around 70%, and results were available hours after the game ended, limiting their value for live commentary and in-game strategy.
Using Moondream fine-tuned with reinforcement learning on NBA broadcast footage, the system identifies the ball carrier in each frame with an F1 score of 0.79. The model was trained in just 60 steps and eliminates the false positives that plagued the base model, dropping from 61 false detections to 2 per game.
- F1 score improved from 0.28 to 0.79, outperforming GPT-5.4 (0.53)
- False positives reduced from 61 to 2 per game
- Results available in real time instead of hours post-game
- Eliminated manual frame-by-frame review for possession tracking
Complete Vision AI Stack
This solution uses Moondream's integrated stack from model training through production deployment. Every layer is designed to work together, so you go from problem to deployed system without stitching together tools from different vendors.
AI Model Layer
Base Model
Moondream 3
Fine-Tuning
Reinforcement Learning via Lens
Production Model
Moondream 3 (fine-tuned)
Deployment Layer
Inference Engine
Photon
Target Hardware
NVIDIA L40
Deployment
Cloud
Training Method
RL
Training Steps
60
Task Type
detect
F1
28.3% → 78.8%
Ready to build your solution?
Talk to our team about how Moondream can solve your specific vision AI challenge, from model training through production deployment.