SDSTrack

This is an official release of the CVPR 2024 paper: SDSTrack: Self-Distillation Symmetric Adapter Learning for Multi-Modal Visual Object Tracking.

Models & Raw Results(Google Driver) Models & Raw Results(Baidu Driver: qolo)

Pipeline

News

[Mar 26, 2024]

✅ We release codes, models and raw results.

[Feb 27, 2024]

✅ SDSTrack is accepted to CVPR2024.

Highlights

💡 SDSTrack uses lightweight adapters for parameter-efficient fine-tuning and can be used for various modality combinations.

💡 We employ a symmetric design, treating all modalities equally, thereby reducing modality dependencies.

💡 We significantly improve the robustness of trackers through complementary masked patch distillation.

💡 SDSTrack performs excellently even in extreme scenarios, such as modality absence and severe occlusion.

Usage

Install the environment

Create and activate a conda environment:

conda create -n sdstrack python=3.8
conda activate sdstrack

Install the required packages:

bash install_sdstrack.sh

Data Preparation

Put the training datasets in ./data/. It should look like:

$<PROJECT_ROOT>
-- data
    -- depthtrack/train
        |-- adapter02_indoor
        |-- bag03_indoor
        ...
    -- lasher/trainingset
        |-- 1boygo
        |-- 1handsth
        ...
    -- visevent/train
        |-- 00142_tank_outdoor2
        |-- 00143_tank_outdoor2
        ...
        |-- trainlist.txt

Set project paths

Run the following command to set paths:

cd <PROJECT_ROOT>
python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir ./output

You can also modify paths with these two files if needed:

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

Training

Download the pretrained foundation model (OSTrack) and put it under $PROJECT_ROOT$/pretrained. To train RGB-Depth tracking, you can run:

bash train_sdstrack_rgbd.sh

You can also train various modality combinations by bash train_sdstrack_rgbe.sh and bash train_sdstrack_rgbt.sh.

Evaluation

RGB-Depth benchmarks

Two benchmarks (DepthTrack testing set and VOT22_RGBD) are evaluated using VOT-toolkit.
- Put the DepthTrack testing set to ./Depthtrack_workspace/ and rename it sequences.
- Download the VOT22_RGBD testing set, put it to./VOT22RGBD_workspace/ and also rename it sequences.
```
bash eval_rgbd.sh
```
RGB-Thermal benchmarks

Two benchmarks (LasHeR testing set and RGBT234) are evaluated.
```
bash eval_rgbt.sh
```
We refer you to LasHeR Toolkit for LasHeR evaluation and MPR_MSR_Evaluation for RGBT234 evaluation.
RGB-Event benchmarks

One benchmark (VisEvent) is evaluated. Please modify the and in./RGBE_workspace/test_rgbe_mgpus.py, then run:
```
bash eval_rgbe.sh
```

Acknowledgments

Thanks for the OSTrack and PyTracking library, which help us to quickly implement our ideas.
We Thank for the ViPT, which is an excellent and inspiring work.

Citation

If our work is useful for your research, please consider citing:

  @InProceedings{Hou_2024_CVPR,
    author    = {Hou, Xiaojun and Xing, Jiazheng and Qian, Yijie and Guo, Yaowei and Xin, Shuo and Chen, Junhao and Tang, Kai and Wang, Mengmeng and Jiang, Zhengkai and Liu, Liang and Liu, Yong},
    title     = {SDSTrack: Self-Distillation Symmetric Adapter Learning for Multi-Modal Visual Object Tracking},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {26551-26561}
}

hoqolo / SDSTrack

readme