Olympic Boxing Fight Video Dataset for Punch Detection and Classification

This dataset contains synchronized video recordings of Olympic (amateur) boxing fights, collected during an official 2021 boxing league event held in Szczyrk, Poland, featuring schoolboy, junior, and youth categories. It is designed to support research on punch detection, classification, and action recognition in combat sports.

Recording setup. Four GoPro HERO8 cameras were mounted on tripods behind each corner of the ring and recorded in Full HD (1920 × 1080) at 50 fps, providing consistent multi-angle coverage of 22 fights (~4 hours of footage). The recordings are fully synchronized across all four cameras at the frame level (630,965 frames per camera), so identical frame indices correspond to the same temporal moment in every view.

Annotations. A subset of the footage from the two cameras positioned behind the neutral corners (kam2 and kam4) was manually annotated frame-by-frame in CVAT by two domain experts, including a licensed Olympic boxing referee. Each annotation provides a bounding box per frame and one of eight action classes: punch to the head (left/right hand), punch to the torso (left/right hand), block (left/right hand), and missed punch (left/right hand). Annotations are stored in JSON files accompanying the corresponding MP4 recordings; each file also lists the frame numbers explicitly reviewed by the referee, which is essential when building training and evaluation splits (punches may occur in unreviewed frames).

Privacy. Faces of all individuals in the footage were blurred using the MogFace face-detection algorithm, following an interpretation issued by the legal department of the University of Economics in Katowice.

Structure. The archive contains a database directory with per-camera subdirectories (kam1–kam4) holding the MP4 recordings and JSON annotation files, and a scripts directory with a Python visualization script (database_review.py), a demonstration video (database_review.mp4), and a README.md with usage instructions.

Benchmark. The dataset is accompanied by a benchmark of six state-of-the-art object detectors (Cascade R-CNN, Faster R-CNN, RetinaNet, RT-DETR, YOLO11m, YOLOv8m) trained on the annotated portion using a standardized temporal five-fold validation procedure. The full experimental code is available at: github.com

To our knowledge, this is the first publicly available, real-world benchmark for punch detection and classification in Olympic boxing.