beach-volley-vision-models

Beach Volley Vision — Fine-tuned TrackNetV3 (Ball Tracking)

A fine-tuned TrackNetV3 model for tracking the ball in beach volleyball footage. TrackNetV3 is a heatmap-based tracker designed for small, fast-moving sports objects (originally shuttlecocks); this checkpoint adapts it to the beach volleyball domain.

This is a derivative of the upstream TrackNetV3 model, fine-tuned on hand-annotated beach volleyball rally clips. The upstream project is MIT licensed, and this derivative carries the same license forward (see License).

Model details

• Architecture: TrackNetV3 (trajectory-prediction module). This repo contains the fine-tuned TrackNet weights only.
• Base model: qaz812345/TrackNetV3
• Task: per-frame ball localization (heatmap → (x, y) + visibility)
• Domain: beach volleyball match video
• File: tracknet_best.pt (~130 MB)

Note on InpaintNet: the trajectory-rectification (inpainting) module was not retrained. If your pipeline uses it, download InpaintNet_best.pt from the upstream TrackNetV3 repo.

Usage

This checkpoint is consumed by the Beach Volley Vision pipeline: 👉 https://github.com/ddecks/beach-volley-vision

# Download the fine-tuned weights (git-lfs required)
git lfs install
git clone [email protected]:deadfast/beach-volley-vision-models
# or grab the single file:
#   https://huggingface.co/deadfast/beach-volley-vision-models/resolve/main/tracknet_best.pt

# Place it where the pipeline expects it
mkdir -p data/models
cp beach-volley-vision-models/tracknet_best.pt data/models/tracknet_best.pt

See the project repository for the full inference pipeline (detection → tracking → rally/event detection → stats).

Training data

Fine-tuned on hand-annotated beach volleyball rally clips — frames labeled with ball position and visibility, exported to the TrackNet CSV format (Frame,Visibility,X,Y). Example annotation and evaluation data is included in the project repository under data/.

License

MIT — carried forward from the upstream TrackNetV3 project (Copyright (c) 2024 qaz812345). Fine-tuning and beach volleyball adaptation by Devin Decker.

Acknowledgements

• TrackNetV3 by qaz812345 — the base architecture and pretrained weights this model is fine-tuned from.