README.md

🔬 Related Resources These work are parts of our research on Closely-Spaced Infrared Small Target Unmixing For a comprehensive collection of papers, datasets, and resources, visit: 📚 View Awesome-CSIST-Unmixing

This repository contains the official implementation of the following papers:

DISTA-Net: Dynamic Closely-Spaced Infrared Small Target Unmixing
Shengdong Han, Shangdong Yang, Yuxuan Li, Xin Zhang, Xiang Li, Jian Yang, Ming-Ming Cheng, Yimian Dai
ICCV 2025. Paper Link | 中文论文翻译 | 博客解读 | 视频讲解

SeqCSIST: Sequential Closely-Spaced Infrared Small Target Unmixing
Ximeng Zhai, Bohan Xu, Yaohong Chen, Hao Wang, Kehua Guo, Yimian Dai
IEEE TGRS 2025. Paper Link | 博客解读 | 视频讲解

Beyond Unfolding: 60× Faster One-Stage Unmixing for Closely-Spaced Infrared Small Targets
Ximeng Zhai, Zheng Wang, Zhu Liu, Yaohong Chen, Hao Wang, Kehua Guo, Ming-Ming Cheng Yimian Dai
arXiv. Paper Link

📘 Introduction

An open-source ecosystem for the unmixing of closely-spaced infrared small targets including:

• CSIST-100K, a publicly available benchmark dataset for single-frame CSIST Umixing;

• SeqCSIST, a publicly available benchmark dataset specifically designed for multi-frame CSIST Umixing.

• CSO-mAP, a custom evaluation metric for sub-pixel detection;

• GrokCSO, an open-source toolkit featuring DISTA-Net and other models.

---

🗂 Datasets

CSIST-100K Dataset

A synthetic dataset for multi-target sub-pixel resolution analysis under diffraction-limited conditions. Download: Baidu Pan / OneDrive.

Parameter	Value/Range
Imaging Size	11×11 pixels
$σ_{PSF}$	0.5 pixel
Targets per Image	1–5 (random)
Intensity Range	220–250 units (uniform)
Spatial Constraints	Sub-pixel coordinates within a pixel + 0.52 Rayleigh unit separation

SeqCSIST Dataset

A synthetic dataset specifically designed for multi-frame CSIST Unmixing, consisting of 100,000 frames organized into 5,000 random trajectories. Download: Baidu Pan

🏗 Networks

Architecture of the proposed DISTA-Net. The overall framework consists of multiple cascaded stages. Each stage contains three main components: a dual-branch dynamic transform module ($\mathcal{F}^{(k)}$) for feature extraction, a dynamic threshold module ($\Theta^{(k)}$) for feature refinement, and an inverse transform module ($\tilde{\mathcal{F}}^{(k)}$) for reconstruction.

Architecture of the proposed DeRefNet. The overall framework consists of three main modules: a sparsity-driven feature extraction module for effective CSIST feature extraction through nonlinear learnable and sparsifying transforms, a positional encoding module for temporal information enhancement to enable finer sub-pixel target localization, and a temporal deformable feature alignment (TDFA) module for dynamic reference-based refinement through multi-frame deformable alignment at the feature level.

Overview of the FOCUS framework and its physics-constrained optimization strategy. (a-b) The architecture utilizes a coarse-to-fine flow via DMM and SMM modules, incorporating Isomorphic Source-Generation Units (ISGU) that bifurcate intensity and mask paths to ensure discrete source recovery. (c-e) Comparison of label characteristics and predictive evolution between standard super-resolution (SR) and the proposed unmixing paradigm. The evolution under joint L_MSE, L_Sparsity, and L_Energy constraints, visualized through the optimization vectorgraph, illustrates the competitive equilibrium achieved to maintain structural sparsity and flux conservation for high-fidelity signal reconstruction.

📈 Comparison with state-of-the-art methods

📘GrokCSO Instructions

🛠️Environment Preparation

Installation

$ conda create --name grokcso python=3.9 
$ source activate grokcso

Step 1: Install PyTorch

# CUDA 12.1  
conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=12.1 -c pytorch -c nvidia  

Step 2: Install OpenMMLab 2.x Codebases

$ pip install -U openmim
$ pip install mmcv==2.1.0 -f https://download.openmmlab.com/mmcv/dist/cu121/torch2.1/index.html

$ pip install mmdet

Step 3: Install grokcso

$ git clone https://github.com/GrokCV/GrokCSO.git
$ cd grokcso
$ python setup.py develop

🚀Run Script

✨Train a model：

# c = 3  
$ CUDA_VISIBLE_DEVICES=1 python tools/train.py --config configs/Agrok/dista.py  
  
# c = 5  
$ CUDA_VISIBLE_DEVICES=1 python tools/train.py --config configs/c_5/dista.py  
  
# c = 7  
$ CUDA_VISIBLE_DEVICES=1 python tools/train.py --config configs/c_7/dista.py   

✨Test a model：

# c = 3  
$ CUDA_VISIBLE_DEVICES=1 python tools/test.py --config configs/fdist/dista.py --checkpoint /pth/dista/epoch_47.pth --work-dir work_dir/dista
  
# c = 5  
$ CUDA_VISIBLE_DEVICES=1 python tools/test.py --config configs/c_5/dista.py --checkpoint /pth/dista/c_5/epoch_105.pth --work-dir work_dir/dista/c_5
  
# c = 7  
$ CUDA_VISIBLE_DEVICES=1 python tools/test.py --config configs/c_7/dista.py --checkpoint /pth/dista/c_7/epoch_246.pth --work-dir work_dir/dista/c_7

🎁Citation

@inproceedings{han2025dista,
  title={{DISTA-Net}: Dynamic Closely-Spaced Infrared Small Target Unmixing},
  author={Han, Shengdong and Yang, Shangdong and Li, Yuxuan and Zhang, Xin and Li, Xiang and Yang, Jian and Cheng, Ming-Ming and Dai, Yimian},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={14655--14664},
  year={2025}
}

@article{zhai2025seqcsist,
  title={{SeqCSIST}: Sequential Closely-Spaced Infrared Small Target Unmixing},
  author={Zhai, Ximeng and Xu, Bohan and Chen, Yaohong and Wang, Hao and Guo, Kehua and Dai, Yimian},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2025},
  publisher={IEEE}
}