AISoccer Coach - Computer Vision Developer
- Present
- Built a pipeline that groups tracked soccer players by team and identifies goalkeepers and referees. Compared 15 MobileNetV2 and ResNet50 models, with the best reaching 97.31% mean per-video-group clustering accuracy on a development benchmark of 5,581 crops.
- Developed a video pipeline that recognizes 14 player actions and eight ball events using a fine-tuned Perception Encoder and temporal convolutions. The best ball-event model reached 76.2% macro-F1 on 180 labeled development clips.
- Developed a pipeline that estimates a top-down 2D ball path from single-camera soccer footage by combining trajectory analysis, contact classification with a vision-language model, and player tracking.