FDTA: From Detection to Association

Official model weights for the paper "From Detection to Association: Learning Discriminative Object Embeddings for Multi-Object Tracking" (CVPR 2026).

TL;DR. We reveal that DETR-based end-to-end MOT suffers from overly similar object embeddings. FDTA explicitly enhances discriminativeness in this paradigm.

Available Checkpoints

File	Dataset
dancetrack.pth	DanceTrack
sportsmot.pth	SportsMOT

Main Results

DanceTrack

Training Data	HOTA	IDF1	AssA	MOTA	DetA
train	71.7	77.2	63.5	91.3	81.0
train+val	74.4	80.0	67.0	92.2	82.7

SportsMOT

Training Data	HOTA	IDF1	AssA	MOTA	DetA
train	74.2	78.5	65.5	93.0	84.1

BFT

Training Data	HOTA	IDF1	AssA	MOTA	DetA
train	72.2	84.2	74.5	78.2	70.1

Usage

Download Checkpoints

from huggingface_hub import hf_hub_download

# Download the DanceTrack checkpoint
ckpt_path = hf_hub_download(
    repo_id="Spongebobbbbbbbb/FDTA",
    filename="dancetrack.pth",
    local_dir="./checkpoints/"
)

For full training and evaluation instructions, please refer to the GitHub repository.

Citation

@article{shao2025fdta,
  title={From Detection to Association: Learning Discriminative Object Embeddings for Multi-Object Tracking},
  author={Shao, Yuqing and Yang, Yuchen and Yu, Rui and Li, Weilong and Guo, Xu and Yan, Huaicheng and Wang, Wei and Sun, Xiao},
  journal={arXiv preprint arXiv:2512.02392},
  year={2025}
}

License

This project is released under the MIT License.