TensorISTD

TensorISTD is an open-source MATLAB toolbox designed for optimization-based Infrared Small Target Detection (ISTD) research and evaluation.

It provides a unified, streamlined evaluation pipeline covering:

• Single-frame / Multi-frame tasks
• Matrix / Tensor representations
• Mainstream evaluation metrics (SCRG, BSF, CG, and 3D-ROC series)

Objectives:

1. Enable rapid prototyping and verification of optimization-based detection methods.
2. Establish a standardized benchmark for fair comparison and reproducibility.
3. Encourage community contributions to enrich the algorithm library.

---

Table of Contents

• TensorISTD
  • Requirement
  • Usage Instruction
    • Get Results
    • Evaluation
    • Draw Visualization Images
  • Evaluation Table
  • Citation
  • References
  • Acknowledgements

Note:

1. This benchmark also includes the code of our profound and powerful baseline, STPA-FCTN, which can be found in TensorISTD/algorithms/STPA_FCTN.
2. This repository will be updated regularly, so stay tuned for improvements and new features!
3. If you would like to contribute, please contact us!

---

Requirement

Matlab 2021b or higher.

Usage Instruction

📂 Document Structure

TensorISTD/
├── algorithms/               # Core detection algorithms
│   ├── 4D_ISTD/              # [1]
│   ├── LogTFNN/              # [2]
│   ├── NPSTT/                # [3]
│   ├── PSTNN/                # [4]
│   └── STPA_FCTN/            # Ours
├── all_results/              # Raw detection outputs
│   └── sequence1/            # Sample sequence
│       └── STPA_FCTN/        # Algorithm-specific results
├── fig_results/              # Visualization outputs
│   └── sequence1/            # Sequence-specific figures
├── mat_results/              # Performance metrics
│   ├── curve_results/        # .mat documents
│   └── index_results/        # Metrics records(3D-ROC,SCR, etc.)
├── time_results/             # Runtime logs
│    └── STPA-FCTN.txt        # Time consumption record
├── utils/                    # Support utilities
│   ├── analyse_pts.m         # Point analysis script
│   ├── curves_draw.m         # Visualization generator
│   ├── get_algo_result.m     # Algorithm output collector
│   ├── get_curves.m          # Metrics calculator
│   ├── measure_calculator.m  # SCR/CG/BSF/BSR calculator
│   └── pt_nms.m              # Non-maximum suppression
├── 3D_Visualization          
├── LICENSE                   # open-source license
├── README.md
├── color1.mat           
└── evaluation.m              

🚀 Get Results

Main command

evaluation.m

Use all the evaluation algorithms to get the result plots for all the evaluation datasets present in . /result in the mat file.

1. Inititalization

Select the desired algorithm name and datasets：

eval_algo_names = ...
    {
     'STPA_FCTN' %'TT','TR','LogTFNN','NPSTT','PSTNN','RIPT','STT'
    };

eval_data_names = ...
    {
     'sequence1' 
    };

Fill in the following configuration：

img_types = {'*.jpg', '*.bmp', '*.png'}; 
algo_base_path = '.\algorithms\'; 
data_base_path = '.\dataset\data\';
res_base_path =  '.\all_result\'; 
time_path = '.\time_results\';

2. Choose single-frame or multi-frame algorithm：

Single-frame (PSTNN, etc.)

get_algo_result(eval_algo_names, eval_data_names, ...
     img_types, algo_base_path, data_base_path, res_base_path, time_path );

Multi-frame (STT, NPSTT, 4-D-TT/TR, STPA-FCTN, etc.)

get_algo_result_multiframe(eval_algo_names, eval_data_names, ...
     img_types, algo_base_path, data_base_path, res_base_path, time_path );

💡 Note

If the metrics are calculated directly from the existing test image, then comment out this section and go directly to Evaluation.

📈 Evaluation

In this repo, we include the following metrics:

✅ SCRG, ✅ CG, ✅ BSF,

✅ Diverse AUC Analysis ([6]): $AUC_{(FPR,TPR)}$, $AUC_{(\tau,TPR)}$, $AUC_{(\tau,FPR)}$, $AUC_{ODP}$, $AUC_{SNPR}$, $AUC_{TD}$, $AUC_{BS}$, $AUC_{TDBS}$

1. Calculate the corresponding .mat file based on Get Results or the existing result plots with the target coordinates of the dataset and store it in the curve index folder.

get_curves(eval_algo_names, eval_data_names, thres_num, radius, res_base_path, ...
     mat_base_path, txt_base_path, mask_base_path, preimg_type);

This step will combine the result plots from step 1 and . GT under /dataset/ann/ to get the roc sequence results

• Related Configurations：

%% evaluation.m
mat_base_path = '.\mat_results\'; 
txt_base_path =  '.\dataset\anno\'; 
preimg_type = '*.png';

2. Calculate the Multi-perspective AUC Analysis:

curves_drawer(1, eval_algo_names, eval_data_names, figure_base_path, mat_base_path, x_axis_ratio, FPR_thres);

3. Calculate the SCRG gain, CG, BSF metrics:

measure_calculator(eval_algo_names, eval_data_names, data_base_path, res_base_path, ...
    mat_base_path, txt_base_path, img_types, preimg_type);

4. Draw the 3D-ROC figures

• 📂The evaluation result will be saved in index_results:

├──./mat_result/
│    ├── curve_results
│    │    ├── sequence1_STPA-FCTN.mat
│    │    ├── ...
│    ├── index_results
│    │    ├── sequence1.txt
│    │    ├── ...

• The 3D-ROC image can be obtained during the execution of step 3.

  • Related Configurations：

  %% evaluation.m
  figure_base_path = '.\fig_results\';
  x_axis_ratio = 1e-4;
  FPR_thres = 1;

  %% utils\curves_drawer.m
  % Line Color
      color_map = [ ...
  %          55/255   126/255  184/255;  % Blue
  %          77/255   175/255  74/255;   % Green
  %          228/255  26/255   28/255;   % Red
  %          255/255  217/255  47/255;   % Yellow
  ];
  % Line Type
  LineType = {':' }; %'-.'

  • 📂 The evaluation 3D-ROC result has the following structure:

  ├──./fig_results/
  │    ├── sequence1
  │    │    ├── SOTA_1
  │    │    ├── SOTA_2
  │    │    ├── SOTA_3
  │    │    ├── SOTA_4
  │    ├── sequence2
  │    │    ├── SOTA_1
  │    │    ├── ...
  │    ├── ...
  

  • The following figures are 3D-ROC results of STPA-FCTN in seq 1.

  

  • Comparison of multiple algorithms.

  

🎨 Draw Visualization Images

• This following script provides a standardized pipeline for generating publication-quality 3D visualizations from 2D images.

  • Key Configuration Parameters
  1. Figure Window

  figure('Units', 'pixels', 'Position', [100 100 703 641]);

  2. Axis Configuration

  xticks(0:50:250);          
  xtickformat('%d');          
  xlim([0 250]);             
  xtickangle(0);         
  ...

  3. View Perspective

  view(-37.5,30);              % Set 3D view perspective (azimuth -37.5°, elevation 30°)

  4. Color Configuration

  The colormap is provided as color1.mat.

  • 3D Visualization

  

Evaluation Table

Method	Averaged Values
	SCRG	BSF	CG	$AUC_{FPR,TPR}$	$AUC_{\tau,TPR}$	$AUC_{\tau,FPR}$	$AUC_{ODP}$	$AUC_{SNPR}$	$AUC_{TD}$	$AUC_{BS}$	$AUC_{TDBS}$	Time(s)
PSTNN	Inf	Inf	1.4788	0.9429	0.7011	6.1038e-3	1.6249	1.2251e2	1.6310	0.9238	0.6950	0.1606
LogTFNN	4.5912	5.1813	1.0208	0.9756	0.5694	6.5000e-2	1.5385	0.9629e2	1.5450	0.9629	0.5629	1.3208
NPSTT	Inf	Inf	1.8846	0.9699	0.7824	5.1423e-3	1.7517	1.5926e2	1.7523	0.9649	0.7818	5.1299
4-D TT	89.171	41.837	2.3124	0.9921	0.9277	5.1489e-3	1.9147	1.8021e2	1.9199	0.9869	0.9226	0.8694
4-D TR	100.62	49.084	2.2463	0.9976	0.9374	5.0947e-3	1.9300	1.8409e2	1.9351	0.9926	0.9324	2.0373
STPA-FCTN	149.26	68.905	2.3292	0.9996	0.9581	5.0778e-3	1.9527	1.8893e2	1.9579	0.9945	0.9530	0.9926

💡 Note

1. In multi-frame evaluation, an "Inf" value does not necessarily indicate that all frames produced infinite results. This can occur if even a single frame yields an "Inf" value. Therefore, for multi-frame scenarios, we recommend using CG or 3-D ROC metrics for more robust and reliable assessment.

2. We adopted sequences from [1], and we also provide our data collection [2].

Citation

If you find the code useful, please consider citing our paper using the following BibTeX entry.

@ARTICLE{11333368,
  author={Wu, Fengyi and Chen, Siyu and Liu, Simin and Tao, Bingjie and Luo, Junhai and Peng, Zhenming},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={A Comprehensive Benchmark for Spatiotemporal Tensor-Based Infrared Small Target Detection}, 
  year={2026},
  volume={64},
  number={},
  pages={1-20},
  keywords={Tensors;Benchmark testing;Three-dimensional displays;Correlation;Object detection;Transformers;Computational modeling;Accuracy;Systematics;Clutter;Benchmark;fully connected tensor network completion;infrared small target detection (ISTD)},
  doi={10.1109/TGRS.2026.3651631}}

References

[1] F. Wu, H. Yu, A. Liu, J. Luo, and Z. Peng, “Infrared small target detection using spatiotemporal 4-d tensor train and ring unfolding,” IEEE Trans. Geosci. Remote Sens., vol. 61, pp. 1–22, 2023.

[2] X. Kong, C. Yang, S. Cao, C. Li and Z. Peng, "Infrared Small Target Detection via Nonconvex Tensor Fibered Rank Approximation," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-21, 2022.

[3] G. Wang, B. Tao, X. Kong, and Z. Peng, “Infrared small target detection using nonoverlapping patch spatial–temporal tensor factorization with capped nuclear norm regularization,” IEEE Trans. Geosci. Remote Sens., vol. 60,pp. 1–17, 2021.

[4] L. Zhang and Z. Peng, “Infrared small target detection based on partial sum of the tensor nuclear norm,” Remote Sens., vol. 11, no. 4, p. 382, 2019.

[5] C. -I. Chang, "An Effective Evaluation Tool for Hyperspectral Target Detection: 3D Receiver Operating Characteristic Curve Analysis," in IEEE Transactions on Geoscience and Remote Sensing, vol. 59, no. 6, pp. 5131-5153, June 2021,

Note: If you used the above codes, please cite the relevant paper.

Acknowledgements

Despite the organizers of this repo, we would like to thank the former contributors--Tianfang Zhang, Jian Li, Yuelu Wei, Guanghui Wang, Xuan Kong, Haiyang Yi, Ruochen Qie, Hang Yu, Anran Liu, Simin Liu, and Zhenming Peng--for this Toolbox.