bambezius/soccernet-v3-ball-tracking

SoccerNet-v3 ball tracking — private transfer

Private transfer copy for the account owner's authorized SoccerNet project. Original source: https://github.com/SoccerNet/SoccerNet-v3 Original SoccerNet access and usage terms continue to apply; this copy grants no additional rights and should not be made public.

Includes 400 original per-game Frames-v3.zip archives, Labels-v3.json files, and the canonical ball manifests from https://github.com/anthonyceponis/glitch. Official splits: 25,311 train, 3,945 validation, 4,730 test frames. No checkpoints, credentials, or prediction artifacts are included.

Restore into a clone of the project

Authenticate with a Hugging Face token that can read this private repository. From the project root, run (replace bambezius/soccernet-v3-ball-tracking with this repository's ID):

hf download bambezius/soccernet-v3-ball-tracking --repo-type dataset --local-dir . --include 'data/**'
PYTHONPATH=src python scripts/import_soccernet_v3.py
python scripts/make_yolo_training_subset.py --train-manifest data/processed/soccernet-v3/train.csv --val-manifest data/processed/soccernet-v3/valid.csv --output-dir artifacts/yolo11-soccernet-data --train-count all --val-count all --seed 0

The importer extracts frames and rebuilds the canonical manifests. The second script builds machine-local YOLO symlinks. Install project dependencies first; use a fresh output directory for the YOLO preparation script. Hugging Face handles resumable transfers and content hashes. transfer-manifest.json records the uploaded source paths and sizes. Reserve room for both archives and extracted frames. Checkpoints are transferred separately, for example via W&B.