Quick Start

The soure code of the paper "Multi-domain Universal Representation Learning for Hyperspectral Object Tracking".

1. Environment Setting

The environment configuration follows https://github.com/jiawen-zhu/ViPT.

2. Hyperspectral Video Dataset

• The HOT2023, HOT2020, and HOT2022 datasets are from "https://www.hsitracking.com/".
• The IMEC25 dataset is from paper "Histograms of oriented mosaic gradients for snapshot spectral image description".

3. Path Setting

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

4. First-satge train in HOT2023

(a) Download pretrained model and put in the folder "pretrained_model", which is available in
- https://pan.baidu.com/s/1qRuCKQ2hhE5-MhrkeLiEQA - Access code: 2025

(b) Change the path of training data in lib/train/base_functions.py (Line 100: settings.env.hsi_dir='/data/XXX/XX')

(c) Run: python tracking/train.py --script vipt --config deep_all --save_dir ./output

5. Second-satge train in HOT2023

(a) Use the model trained in first stage and put in the folder "pretrained_model", which is available in
- https://pan.baidu.com/s/1WJLo72hwzr6y_BtjFFp-Dg - Access code: 2025

(b) Change the path of training data in lib/train/base_functions.py (Line 100: settings.env.hsi_dir='/data/XXX/XX')

(c) Fix all parameter, only train the domain adapter in each hyperspectral domain.

(d) Run: python tracking/train.py --script vipt --config deep_all --save_dir ./output

6. Test

(a) Download testing model in
- https://pan.baidu.com/s/1WJLo72hwzr6y_BtjFFp-Dg - Access code: 2025

(b) Put the testing model in the folder "final_model".

(c) Run in HOT2023:

VIS domain: python test_hsi_mgpus_all.py --dataset_name HOT23TEST --data_path /data/lizf/HOT/Whispers2023/validation/HSI-VIS --model_path final_model_path_HOT2023
NIR domain: python test_hsi_mgpus_all.py --dataset_name HOT23TEST --data_path /data/lizf/HOT/Whispers2023/validation/HSI-NIR --model_path final_model_path_HOT2023
RedNIR domain: python test_hsi_mgpus_all.py --dataset_name HOT23TEST --data_path /data/lizf/HOT/Whispers2023/validation/HSI-RedNIR --model_path final_model_path_HOT2023

(d) Run in HOT2020 and HOT2022 (use the trained model in HOT2023):

VIS domain: python test_hsi_mgpus_all.py --dataset_name HOT23TEST --data_path /data/lizf/HOT/Whispers2023/validation/HSI-VIS --model_path final_model_path_HOT2023

(e) Run in IMEC25 (fine-tune the parameter of NIR adapter in IMEC25):

NIR domain: python test_hsi_mgpus_all.py --dataset_name HOT23TEST --data_path /data/lizf/HOT/Whispers2023/validation/HSI-VIS --model_path final_model_path_IMEC25

7. Cite

@article{LI2025111389,
title = {Multi-domain universal representation learning for hyperspectral object tracking},
author = {Zhuanfeng Li and Fengchao Xiong and Jianfeng Lu and Jing Wang and Diqi Chen and Jun Zhou and Yuntao Qian},
journal = {Pattern Recognition},
volume = {162},
pages = {111389},
year = {2025},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2025.111389},
}

8. Concat

• lizhuanfeng@njust.edu.cn;
• If you have any questions, just contact me.