besein-datasets — Besein Soccer Analytics data backup (private)

Snapshot taken 2026-09-06 from the Mac (primary data store). Mirrors the weights repo innovazets/besein-models.

Contents

Path	Content	Size
public_ball_parsed/	Parsed public ball dataset (1,237 frames, real images+labels) — ball class diversity for ball models	~270 MB
labels_archive/sportsmot_gt_yolo_labels.tar.gz	Real-file labels only: 18,834 clean SportsMOT GT labels (13,664 train / 5,170 val), 4-class mobadam schema + yaml	3.2 MB
labels_archive/sportsmot_gt_labels.tar.gz	Older SportsMOT GT label export	5.3 MB
labels_archive/player_v14_gt_labels.tar.gz	Half-A relay selection (v14/v15 source) — mostly symlinks into gt_yolo, tar is near-empty by design	4 KB

What is intentionally NOT here (and where it lives)

• SportsMOT image corpus (train_sets/public_new/sportsmot, ~35 GB): real jpgs for train/val/test (45/45/150 seqs). Re-downloadable from the public SportsMOT source; labels in this repo + mirror on the Mac. Do NOT mirror this upstream data to HF.
• Assembled training datasets (ds17mc 14,756 train frames, ds17mc_clips 1,886 clip frames, keremberke_autolabel 812 frames): assembled/auto-labeled artifacts that live on the RunPod A40 volume /workspace/basein/train_sets (pod currently STOPPED, volume kept — nothing lost). Rebuild from this repo + the label tars + the image corpus via the scripts in git (scripts/pod_queues/build_v17_data.sh, autolabel_clips_v171.py, autolabel_dir.py, sample_sportsmot_test.py).
• Bake-off holdout (v16val, 2,585 frames): derived by copying the val subset of the corpus + labels — reproduce from sportsmot_gt_yolo val + mirror.

Rebuild recipe (any machine)

1. Fetch SportsMOT corpus (upstream) → train_sets/public_new/sportsmot.
2. Extract sportsmot_gt_yolo_labels.tar.gz into train_sets/public_new/.
3. Run the build/autolabel scripts from the git repo (pedrobesein/beseinvideotraining).
4. Fine-tune from innovazets/besein-models/player_v18.pt (or mobadam).