🌊 [CVPR 2026 Oral] SEATrack: Simple, Efficient, and Adaptive Multimodal Tracker

📄 Paper  |  📦 Models & Results: Google Drive  /  Baidu Drive /  Hugging Face

---

🔥 News

• [May 1, 2026] The models, raw results, and training logs are now available on Hugging Face 🤗.
• [Apr 13, 2026] Code, models, and raw results are released.

---

🧠 Introduction

• 🌊 A simple unified multimodal tracking framework for RGB-T, RGB-D, and RGB-E tasks.
• 🚀 Achieves strong performance across multiple multimodal benchmarks.
• ⚡ Highly efficient: only 0.6M trainable parameters and 63.5 FPS.
• 🔍 Highlights the importance of cross-modal alignment in multimodal tracking.

---

📊 Results

Overall Performance

Visualization

---

⚙️ Usage

🔧 Installation

conda env create -f environment.yaml
conda activate seatrack

---

📂 Data Preparation

• LasHeR
• RGBT234 (qvsq)
• DepthTrack
• VOT22-RGBD
• VisEvent

Organize datasets as follows:

-- <DATA_PATH>
    -- DepthTrack/trainingset
        |-- adapter02_indoor
        |-- bag03_indoor
        |-- bag04_indoor
        ...
    -- LasHeR/trainingset
        |-- 1boygo
        |-- 1handsth
        ...
    -- VisEvent/trainingset
        |-- 00142_tank_outdoor2
        |-- 00143_tank_outdoor2
        ...

---

🛠 Path Setting

cd <PATH_TO_SEATRACK>
python tracking/create_default_local_file.py \
  --workspace_dir . \
  --data_dir <DATA_PATH> \
  --save_dir ./output

Or manually modify:

./lib/train/admin/local.py # paths for training
./lib/test/evaluation/local.py # paths for testing

---

🏋️ Training

Download pretrained OSTrack and place it under:

./pretrained/vitb_256_mae_32x4_ep300
./pretrained/vitb_256_mae_ce_32x4_ep300

Then run:

bash train.sh

---

🧪 Testing

Modify checkpoint in:

./lib/test/parameter/seatrack.py

RGB-D (DepthTrack & VOT22-RGBD)

Place datasets and the provided list.txt into:

./Depthtrack_workspace/sequences
./VOT22RGBD_workspace/sequences

Modify paths in

./Depthtrack_workspace/trackers.ini
./VOT22RGBD_workspace/trackers.ini

Run evaluation with VOT Toolkit:

bash eval_rgbd.sh

---

RGB-T (LasHeR & RGBT234)

Modify <DATASET_PATH> and <SAVE_PATH> in:

./RGBT_workspace/test_rgbt_mgpus.py

Then run:

bash eval_rgbt.sh

Evaluation tools:

• LasHeR Toolkit
• RGB-T Toolkit

---

RGB-E (VisEvent)

Modify <DATASET_PATH> and <SAVE_PATH> in:

./RGBE_workspace/test_rgbe_mgpus.py

Run:

bash eval_rgbe.sh

Evaluation:

• VisEvent Benchmark

---

📖 Citation

If you find this work helpful, please consider citing:

@misc{su2026seatracksimpleefficientadaptive,
      title={SEATrack: Simple, Efficient, and Adaptive Multimodal Tracker}, 
      author={Junbin Su and Ziteng Xue and Shihui Zhang and Kun Chen and Weiming Hu and Zhipeng Zhang},
      year={2026},
      eprint={2604.12502},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.12502}, 
}

---

🙏 Acknowledgment

SEATrack uses code from a few open source repositories. Without the efforts of these folks (and their willingness to release their implementations), SEATrack would not be possible. We thank these authors for their efforts!

• ViPT
• PyTracking
• LoRA

---

📬 Contact

If you have any questions, feel free to contact:

📧 binbing2024@outlook.com