Overview

This project presents a video analysis pipeline for recognizing player actions and ball events in soccer footage. Given frame-aligned player and ball tracks, the pipeline extracts short clips centered on each tracked subject. Separate classifiers combine features from a fine-tuned Perception Encoder with temporal convolutions to recognize 14 player actions and eight ball events. Predictions from overlapping windows are aggregated and overlaid on the source video to show when actions occur and which tracks they are associated with. Grad-CAM was used to inspect confident player-classification errors across checkpoints. On a development benchmark of 180 labeled ball clips, the best-performing configuration by macro-F1 achieved 76.2%.

Demonstration

Three saved outputs show the pipeline's predictions overlaid on the original match footage. Each clip keeps its original slow-motion playback.

Full Case Study Coming Soon

I’ll share the experiments and decisions behind this demo here soon.