When Fish Look Alike: Tracking Identities with Dual-branch Elasticity

The official implementation of the paper：

When Fish Look Alike: Tracking Identities with Dual-branch Elasticity
Vran Lee, Xin Liu, Yijie Wei, Yeqiang Liu, Hwa Liang Leo, Zhenbo Li* [Project] [Paper] [Code]

---

Contact: vranlee86@gmail.com. Any questions or discussion are welcome!

If like this work, a star 🌟 would be much appreciated!

---

🚀 Updates

• [2026.07] MFT-Edge Testset Eval has been released on Codabench!
• [2026.07] Updates Repo.

🏆 Key Contributions

• Dual-Branch Elastic Framework: We propose TIDE, a highly efficient JDE framework that effectively resolves the computational bottlenecks of tracking dense, homogeneous targets. It provides a scalable dual-branch design to accommodate diverse hardware constraints.
• Minimalist Geometric Association: We introduce AGCIoU, a geometric association metric that maintains robust ID consistency under severe non-rigid deformations and occlusions, completely avoiding the substantial overhead of heavy appearance models.
• Superior Accuracy-Efficiency Balance: Extensive evaluations demonstrate the framework's exceptional accuracy-efficiency trade-off on the MFT-Edge benchmark. The lightweight TIDE-L achieves a competitive HOTA of 28.43 while reducing computational cost by 38.7-fold compared to standard heavy trackers, directly proving its viability for industrial edge deployment.

📊 Tracking Performance

State-of-the-Art Comparison on MFT-Edge Dataset

Method	Params ↓	FLOPs ↓	HOTA ↑	IDF1 ↑	MOTA ↑	IDs ↓
SORT	99.00M	793.21G	22.73	23.91	48.67	2599
ByteTrack	99.00M	793.21G	19.18	19.37	40.17	2325
OC-SORT	99.00M	793.21G	22.99	24.14	48.44	2674
FairMOT	16.55M	72.93G	27.26	29.68	60.74	2456
CMFTNet	45.08M	137.77G	27.08	29.93	61.90	2716
TrackFormer	42.95M	143.43G	26.51	26.73	43.42	899
SU-T	99.00M	793.21G	34.41	40.50	68.52	1902
TIDE-L (Ours)	5.79M	20.47G	28.43	36.29	47.84	574
TIDE-S (Ours)	32.59M	90.13G	29.98	39.01	54.74	908

Click to see the full comparison table

Note: The best results are highlighted in bold, and the second-best results are underlined.

Methods	Params ↓	FLOPs ↓	HOTA ↑	IDF1 ↑	IDP ↑	IDR ↑	DetRe ↑	DetPr ↑	IDs ↓	MOTA ↑	MOTP ↑
SORT	99.00M	793.21G	22.73	23.91	29.09	20.29	44.66	64.03	2599	48.67	72.01
ByteTrack	99.00M	793.21G	19.18	19.37	26.11	15.40	35.66	60.46	2325	40.17	67.99
OC-SORT	99.00M	793.21G	22.99	24.14	29.28	20.54	44.84	63.92	2674	48.44	72.17
HybridSORT	99.00M	793.21G	15.89	17.29	56.77	10.20	11.79	65.58	214	14.23	71.64
QDTrack	57.20M	32.02G	25.27	24.49	27.74	21.93	53.70	67.92	9103	42.81	75.34
FairMOT	16.55M	72.93G	27.26	29.68	36.56	24.98	46.71	68.36	2456	60.74	69.59
CMFTNet	45.08M	137.77G	27.08	29.93	36.35	25.43	47.52	67.93	2716	61.90	69.47
TrackFormer	42.95M	143.43G	26.51	26.73	35.69	21.36	42.04	70.23	899	43.42	76.00
CenterTrack	16.67M	61.36G	22.49	23.39	30.90	18.81	35.11	57.67	1032	26.68	68.48
TransCenter	30.66M	133.09G	27.20	29.48	37.05	24.48	38.22	57.85	597	24.69	73.83
TFMFT	39.93M	215.27G	21.88	26.74	45.55	18.92	29.72	71.54	945	35.65	74.89
SU-T	99.00M	793.21G	34.41	40.50	37.97	43.41	67.45	59.00	1902	68.52	71.81
TIDE-L (Ours)	5.79M	20.47G	28.43	36.29	49.44	28.67	37.34	64.41	574	47.84	67.17
$\Delta$ vs. SU-T	-94.1%	-97.4%	-17.4%	-10.4%	+30.2%	-33.9%	-44.6%	+9.2%	-69.8%	-30.2%	-6.5%
TIDE-S (Ours)	32.59M	90.13G	29.98	39.01	47.87	32.92	42.86	62.32	908	54.74	65.82
$\Delta$ vs. SU-T	-67.1%	-88.6%	-12.9%	-3.7%	+26.1%	-24.2%	-36.5%	+5.6%	-52.3%	-20.1%	-8.3%

🧐 Prerequisites

• CUDA >= 11.0
• Python >= 3.8
• PyTorch >= 1.7.0
• Ubuntu 18.04 or later (Windows is also supported but may require additional setup)

🔧 Installation

• Step.1 Clone this repo.

• Step.2 Install dependencies. We use python=3.8.

  cd {Repo_ROOT}
  conda env create -f requirements.yaml
  conda activate TIDE
  

• Step.3 Perparing datasets.

  e.g. Download MFT_Edge for test (utilize MFT25 or your own datasets as well.)

🏋️ Training Sample

python train.py cmot \
--exp_id YOUR-EXP-NAMES --data_cfg '../src/lib/cfg/mft_edge.json' \
--lr 5e-4 --batch_size 16 --wh_weight 0.5 \
--arch 'tides/tidel' --num_epochs 30 --reid_dim 64

🧪 Testing Sample

python track.py cmot \
--val_MFT_Edge True \
--data_dir /DATASETS/MFT \
--load_model ../exp/cmot/YOUR-EXP-NAMES/model_best.pth \
--arch 'tides/tidel' \ 
--conf_thres 0.4

🔗 Pretrained Models

Our pretrained models can be downloaded from: [BaiduYun: v57h]

🔗 Datasets

MFT_Edge dataset can be downloaded from: [BaiduYun: wfeq]

🙏 Acknowledgements

A large part of the code is borrowed from sompt22, apple, and ultralytics. Thanks for their wonderful works!

📜 Citation

@misc{lee2026fishlookaliketracking,
      title={When Fish Look Alike: Tracking Identities with Dual-branch Elasticity}, 
      author={Vran Lee and Xin Liu and Yijie Wei and Yeqiang Liu and Hwa Liang Leo and Zhenbo Li},
      year={2026},
      eprint={2607.26412},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.26412}, 
}