TrackFish3D Checkpoints

This repository hosts the released checkpoints for TrackFish3D: Self-Supervised 3D Tracking of Schooling Fish from Multi-view Videos.

Code repository:

• https://github.com/pattgene/TrackFish3D

Model repository:

• https://huggingface.co/pattgene/trackfish3d

Included checkpoints

1. trackfish3d-SAM3

• File: trackfish3d-SAM3/best_ckpt.pth
• Setting: SynFish, geometry-only TrackFish3D with SAM3 detections
• Use with: --no_dino
• Reported selection epoch: 396
• Size: 7.6 MB
• SHA-256: 105a66ab7b34fbb7939de3e2b909021386c6c0d3072488f1b3e6651c022bfad3

2. trackfish3d-YOLO26+SAM2

• File: trackfish3d-YOLO26+SAM2/best_ckpt.pth
• Setting: SynFish, geometry-only TrackFish3D with YOLO26+SAM2 detections
• Use with: --no_dino
• Reported selection epoch: 397
• Size: 7.6 MB
• SHA-256: d2f566d9c85c0a7dcaa636f77afdec1aa3f82a3216ae88b1f69b6c96d649c6e4

3. trackfish3d-3D-ZeF

• File: trackfish3d-3D-ZeF/best_ckpt.pth
• Setting: 3D-ZeF, geometry-only TrackFish3D
• Use with: --no_dino
• Reported selection epoch: 392
• Size: 7.6 MB
• SHA-256: e081e148ed481f15a1bfc70901ae86d29b06e0f2b3822980d939fb54c789c536

Usage

Install the TrackFish3D code from:

• https://github.com/pattgene/TrackFish3D

Example inference command:

python infer.py \
  /path/to/best_ckpt.pth \
  /path/to/scene.json \
  /path/to/scene-images \
  --no_dino \
  --output outputs/scene.txt

Example evaluation command:

python evaluate.py \
  --pred outputs/scene.txt \
  --gt /path/to/ground-truth.csv \
  --dist_thr 1.0 \
  --match_pred_range

Use --dist_thr 1.0 for SynFish and --dist_thr 0.5 for 3D-ZeF.

The official 3D-ZeF dataset source is:

• https://huggingface.co/datasets/vapaau/3D-ZeF

Notes

• These checkpoints are the released geometry-only main models. They are not the DINO variant.
• The corresponding public code repository also includes the 3D-SORT baseline and reproduction scripts.
• SynFish data links, project page, and paper links can be added here later once public.