Beyond Family Labels: A Taxonomic Ornstein–Uhlenbeck Prior for Avian 3D Shape Recovery

Checkpoints for the NeurIPS 2026 paper (OpenReview). Code, installation and evaluation: github.com/rocknroll17/BFL.

AniMer+ fine-tuned for 30K steps on CtrlAVES3D with a taxonomic Ornstein–Uhlenbeck prior on the AVES shape codes. The prior is training-only: at inference these are plain AniMer+ models.

File Paper CUB AUC CUB PCK CowBird AUC CowBird PCK CtrlAVES3D AUC md5
bfl.pt Table 1 (seed 0), Table 3, Fig. 4 89.31 57.91 60.91 19.26 93.04 0dbf86fdfd7bdfd79fe9683c474d2bba
ablation/taxonomy_real.pt Fig. 3, Table 2 (real Dtax) 88.92 55.16 61.32 20.25 92.99 752f0694e50705a03451dc8c399bab14
ablation/taxonomy_permuted.pt Fig. 3, Table 2 (permuted Dtax) 88.87 54.90 60.65 18.49 92.95 8c3d5c01c599741d504a4fad801c9f5e

PCK is PCK@0.1, measured with the repository's eval.py on an RTX 4090.

python scripts/download_checkpoints.py      # from the code repository
python eval.py --ckpt checkpoints/bfl.pt --dataset COWBIRD --out-json outputs/bfl/eval_COWBIRD.json

Each file is a torch.save dict with model_state (AniMer+ state dict plus the learned OU scalars ou_prior.*), step and metrics.

Citation

@inproceedings{jeon2026beyond,
  title     = {Beyond Family Labels: A Taxonomic {Ornstein--Uhlenbeck} Prior for Avian {3D} Shape Recovery},
  author    = {Jeon, Yunchan and Kim, SooGon and Kim, Seungwook and Lee, CheolWon and Lee, Jongmin},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support