Every punch,classified.

Upload fight footage. The pipeline runs YOLOv11m detection, YOLOv8m-pose estimation, and AttentionBiLSTM classification across 6 punch types — fully automated.

System modules

Three AI subsystems, one unified pipeline.

Classification Engine

AttentionBiLSTM · 6 classes

Jab
87%
Cross
--
Lead Hook
--
Rear Hook
--
Lead Uppercut
--
Rear Uppercut
--

Neural Telemetry

Real-time pipeline log

Punch Timeline

Per-fighter event stream

F1
F2
0s15s

Scorecards record outcomes.

We record the mechanics.

Wrist velocityElbow angle deltaShoulder rotationTemporal attention30-frame windows
Stage 01 / 03YOLOv11m · best_potential.pt

Person Detection

Fine-tuned on boxing footage, the detector tracks exactly two fighters using Hungarian-algorithm assignment with a 50,000-point inertia bonus that prevents identity switching mid-match.

Bounding-box cropIoU assignment2-fighter lockInertia bonus
Stage 02 / 03YOLOv8m-pose · COCO keypoints

Pose Estimation

17 anatomical landmarks per fighter, per frame. Skeleton integrity validation requires ≥10 visible keypoints before the sequence is passed to the classifier.

17 keypointsElbow angleWrist velocityShoulder width
Stage 03 / 03AttentionBiLSTM · 30-frame windows

Punch Classification

A bidirectional LSTM with temporal self-attention processes 30-frame keypoint sequences (51 dimensions per frame) to classify 6 punch types with real-time confidence scoring.

51-dim input6-class outputTemporal attentionIntensity estimate
System ready

Ready to analyse your footage?

Upload a boxing video and receive a full biomechanical breakdown — punch counts, classifications, and annotated output.

Start analysis