FSH-TAL

Model weights for Automated Temporal Localization of Behavior in Cichlid Fish with Few-Shot Trajectory Tokens (WACV 2027). Code, usage and the full pipeline: https://github.com/bds062/FSH-TAL

Files

Path	Size	What it is
checkpoints/fshtal_release.pt	63 MB	FSH-TAL localizer trained on all 17 recordings
backbone/backbone_release.pt	537 MB	Trokens++ MIL backbone trained on all 17 recordings
checkpoints/fold{0-4}.pt	5 × 63 MB	the paper's 5-fold cross-validation localizers; fold F's never trained on fold F's recordings
features/feats_fold{0-4}/<recording>/feats.npy	3.6 GB	the paper's 5-fold cross-validation feature banks
features/feats_release_all17/<recording>/feats.npy	0.7 GB	the release backbone's feature bank

All inputs come from SAM 3.1 tracks (prompt "fish", streamed at 4 fps in 24 s sessions). Each cross-validation localizer trains on the 13-14 recordings outside its fold and is scored at its final epoch with its fold's bank, giving 41.0 avg mAP (IoU 0.1-0.7) over the 17 held-out recordings.

License

TBD — license to be finalized alongside the associated publication. Contact the authors before reuse.