📘 Introduction

An open-source ecosystem for the unmixing of closely-spaced infrared small targets including: - CSIST-100K, a publicly available benchmark dataset; - CSO-mAP, a custom evaluation metric for sub-pixel detection;

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

Chinese Resources 🇨🇳 📚

• 📄 Chinese Paper Translation: - OneDrive
• 📝 Chinese Article Explanation - Wechat
• 📺 Chinese Video Tutorial - Bilibili

🗂 CSIST-100K Dataset

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

Simulation Parameters

| 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 |

The network

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.

Comparison with state-of-the-art methods

| Method | #P | FLOPs | CSO-mAP | AP-05 | AP-05 | AP-05 | AP-05 | AP-05 | PSNR | SSIM | | :-------- | :-----: | :----: | :-----: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | ISTA | - | - | 7.46 | 0.01 | 0.31 | 2.39 | 9.46 | 25.14 | - | - | | ACTNet | 46.212M | 62.80G | 45.61 | 0.38 | 7.49 | 41.13 | 83.12 | 95.95 | 35.54 | 99.70 | | CTNet | 0.400M | 2.756G | 45.11 | 0.38 | 7.53 | 40.39 | 82.11 | 95.14 | 35.15 | 99.70 | | DCTLSA | 0.865M | 13.69G | 44.51 | 0.39 | 7.35 | 39.35 | 81.15 | 94.34 | 34.63 | 99.65 | | EDSR | 1.552M | 12.04G | 45.32 | 0.33 | 7.07 | 40.58 | 83.24 | 95.41 | 35.37 | 99.71 | | EGASR | 2.897M | 17.73G | 45.51 | 0.42 | 8.03 | 41.32 | 85.71 | 95.08 | 34.57 | 99.66 | | FENet | 0.682M | 5.289G | 45.67 | 0.38 | 7.72 | 41.50 | 83.39 | 95.33 | 35.19 | 99.69 | | RCAN | 1.079M | 8.243G | 45.87 | 0.42 | 7.96 | 41.81 | 83.61 | 95.57 | 35.21 | 99.69 | | RDN | 22.306M | 173.0G | 45.81 | 0.35 | 7.11 | 41.07 | 84.07 | 96.43 | 36.47 | 99.74 | | SAN | 4.442M | 34.05G | 45.95 | 0.36 | 7.35 | 41.17 | 84.32 | 96.57 | 36.50 | 99.74 | | SRCNN | 0.019M | 1.345G | 29.06 | 0.23 | 4.10 | 21.65 | 49.95 | 69.39 | 28.76 | 98.44 | | SRFBN | 0.373M | 3.217G | 46.05 | 0.43 | 9.31 | 42.83 | 83.72 | 94.95 | 34.02 | 99.68 | | HAN | 64.342M | 495.0G | 45.70 | 0.39 | 7.46 | 40.90 | 83.61 | 96.17 | 35.27 | 99.71 | | ISTA-Net | 0.171M | 12.77G | 45.16 | 0.41 | 7.71 | 40.57 | 82.58 | 94.53 | 33.92 | 99.68 | | ISTA-Net+ | 0.337M | 24.33G | 46.06 | 0.42 | 7.66 | 41.58 | 84.46 | 96.17 | 36.09 | 99.72 | | LAMP | 2.126M | 0.278G | 14.22 | 0.05 | 1.11 | 7.31 | 21.56 | 41.06 | 27.83 | 96.89 | | LIHT | 21.10M | 1.358G | 10.35 | 0.06 | 0.92 | 4.99 | 14.74 | 30.05 | 27.51 | 96.42 | | LISTA | 21.10M | 1.358G | 30.13 | 0.25 | 4.13 | 22.29 | 51.18 | 72.82 | 29.89 | 99.12 | | FISTA-Net | 0.074M | 18.96G | 44.66 | 0.45 | 7.68 | 39.74 | 81.24 | 94.19 | 35.75 | 99.67 | | TiLISTA | 2.126M | 0.278G | 14.95 | 0.06 | 1.23 | 7.72 | 22.50 | 46.23 | 27.70 | 97.40 | | ours | 2.179M | 35.10G | 46.74 | 0.38 | 7.54 | 42.44 | 86.18 | 97.14 | 37.87 | 99.79 |

📘GrokCSO Instructions

🛠️Environment Preparation

Installation

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

Step 1: Install PyTorch

```shell

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

```shell $ 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

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

📄Data preparation

👀dataset directory format is as follows：

shell data/ ├── initial_matrix/ │ ├── Q_3.mat # 3×3 sub-pixel division │ ├── Q_5.mat # 5×5 sub-pixel division │ └── Q_7.mat # 7×7 sub-pixel division │ ├── sampling_matrix/ │ ├── a_phi_0_3.mat # 3x sub-pixel division │ ├── a_phi_5.mat # 5x sub-pixel division │ └── a_phi_7.mat # 7x sub-pixel division │ └── cso_data/ # CSO Dateset ├── train/ # (80,000 samples) │ ├── Annotations/ │ │ ├── CSO_00000.xml │ │ ├── ... │ │ └── CSO_79999.xml │ └── cso_img/ # Infrared image files │ ├── image_00000.png │ ├── ... │ └── image_79999.png │ ├── val/ │ ├── Annotations/ # 80000-89999 │ └── cso_img/ │ └── test/ ├── Annotations/ # 90000-99999 └── cso_img/

🚀Run Script

✨Train a model：

```

c = 3

$ CUDAVISIBLEDEVICES=1 python tools/train.py --config configs/Agrok/dista.py

c = 5

$ CUDAVISIBLEDEVICES=1 python tools/train.py --config configs/c_5/dista.py

c = 7

$ CUDAVISIBLEDEVICES=1 python tools/train.py --config configs/c_7/dista.py
```

✨Test a model：

```

c = 3

$ CUDAVISIBLEDEVICES=1 python tools/test.py --config configs/fdist/dista.py --checkpoint /pth/dista/epoch47.pth --work-dir workdir/dista

c = 5

$ CUDAVISIBLEDEVICES=1 python tools/test.py --config configs/c5/dista.py --checkpoint /pth/dista/c5/epoch105.pth --work-dir workdir/dista/c_5

c = 7

$ CUDAVISIBLEDEVICES=1 python tools/test.py --config configs/c7/dista.py --checkpoint /pth/dista/c7/epoch246.pth --work-dir workdir/dista/c_7 ```

🎁Citation

@article{han2025dista, title={DISTA-Net: Dynamic Closely-Spaced Infrared Small Target Unmixing}, author={Han, Shengdong and Yang, Shangdong and Zhang, Xin and Li, Yuxuan and Li, Xiang and Yang, Jian and Cheng, Ming-Ming and Dai, Yimian}, journal={arXiv preprint arXiv:2505.19148}, year={2025} }