Point-to-Mask: From Arbitrary Point Annotations to Mask-Level Infrared Small Target Detection

Weihua Gao; Wenlong Niu; Jie Tang; Man Yang; Jiafeng Zhang; Xiaodong Peng

cs.CVMar 17, 2026

Point-to-Mask: From Arbitrary Point Annotations to Mask-Level Infrared Small Target Detection

Authors: Weihua Gao, Wenlong Niu, Jie Tang, Man Yang, Jiafeng Zhang, Xiaodong Peng

Organizations: Key Laboratory of Electronics and Information Technology for Space Systems, National Space Science Center, Chinese Academy of Sciences, Beijing 100190, China. · School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing 100049, China.

Abstract

Infrared small target detection (IRSTD) methods predominantly formulate the task as pixel-level segmentation, which requires costly dense annotations and is not well suited to tiny targets with weak texture and ambiguous boundaries. To address this issue, we propose Point-to-Mask, a framework that bridges low-cost point supervision and mask-level detection through two components: a Physics-driven Adaptive Mask Generation (PAMG) module that converts point annotations into compact target masks and geometric cues, and a lightweight Radius-aware Point Regression Network (RPR-Net) that reformulates IRSTD as target center localization and effective radius regression using spatiotemporal motion cues. The two modules form a closed loop: PAMG generates pseudo masks and geometric supervision during training, while the geometric predictions of RPR-Net are fed back to PAMG for pixel-level mask recovery during inference. To facilitate systematic evaluation, we further construct SIRSTD-Pixel, a sequential dataset with refined pixel-level annotations. Experiments show that the proposed framework achieves strong pseudo-label quality, high detection accuracy, and efficient inference, approaching full-supervision performance under point-supervised settings with substantially lower annotation cost. Code and datasets will be available at: https://github.com/GaoScience/point-to-mask.

Figures & tables

Fig. 1: Schematic illustration of the proposed framework. It integrates a Data Stream (top) and a Network Stream (bottom). The PAMG algorithm leverages physics priors to evolve random points into pixel-level masks, generating ground truth for the RPR-Net. The RPR-Net then predicts the target geometry parameters from input sequences using a lightweight point regression architecture. This establishes a robust ”Mask-to-Point Supervision” and ”Point-to-Mask Detection” cycle.

Fig. 2: Visual analysis of PAMG robustness. (a) Comparison of masks generated from center ( $\star$ ) and edge ( $\times$ ) seeds, demonstrating that both cover the effective target core. (b) Posterior energy curves showing that despite an initial ”climbing” lag for the edge seed, both trajectories converge to distinct but valid energy peaks after the warm-up phase.

Fig. 3: The graphical user interface of the proposed Label-IRST annotation tool. The interface features a dual-view mechanism designed to reduce visual fatigue and unnecessary cursor movement. The Global View (left) displays the full-resolution raw infrared image for rapid Region of Interest selection. The Detail View (right) automatically renders the selected area in high magnification, allowing users to verify and refine the candidate mask (shown in red). The bottom panel provides controls for annotation parameters and visual enhancement settings.

Fig. 4: Statistical characteristics of the SIRSTD-Pixel dataset. (a) Scale distribution: target areas are mainly concentrated within a small pixel range, consistent with infrared small target characteristics. (b) Saliency distribution: the SCR values span a broad range, reflecting the diversity of target saliency in the dataset.

Method	Setting	SIRSTD	NUDT-SIRST	IRSTD-1K	Mean
MCLC [ 16 ]	Prior	63.08	69.44	71.33	67.95
	Blind	54.42	62.69	66.97	61.36
	$\Delta$	-8.66	-6.75	-4.36	-6.59
MCGC [ 15 ]	Prior	63.58	73.04	63.62	66.75
	Blind	61.01	70.93	63.26	65.07
	$\Delta$	-2.57	-2.11	-0.36	-1.68

TABLE I: Quantitative comparison of pseudo-label generation quality on the SIRSTD, NUDT-SIRST, and IRSTD-1K datasets. All values are reported in mIoU (%).

\Delta

denotes the performance gap between Blind and Prior modes.

Type	Method	Sup.	Out	SIRSTD-Pixel				IRSTD-1K / NUDT-SIRST
Type	Method	Sup.	Out	mIoU $\uparrow$	$P_{d}\uparrow$	$F_{a}\downarrow$	AUC $\uparrow$	mIoU	$P_{d}$	$F_{a}$	AUC
Single-frame	ACM [ 3 ]	Pixel	Mask	43.13	69.04	1.80	0.8452	58.34 / 65.65	94.59 / 97.56	5.23 / 3.05	0.9730 / 0.9878
	AGPCNet [ 40 ]	Pixel	Mask	54.55	71.42	0.33	0.8571	64.66 / 79.78	96.28 / 99.15	3.22 / 2.01	0.9814 / 0.9958
	ALCNet [ 4 ]	Pixel	Mask	45.16	75.50	3.11	0.8775	59.61 / 52.85	96.62 / 98.20	5.12 / 8.57	0.9831 / 0.9910
	DNANet [ 17 ]	Pixel	Mask	55.06	67.81	0.26	0.8390	68.92 / 84.50	94.26 / 98.73	2.58 / 0.82	0.9713 / 0.9936
	ISNet [ 39 ]	Pixel	Mask	47.14	74.19	3.89	0.8710	64.01 / 61.57	95.27 / 98.52	3.74 / 20.5	0.9763 / 0.9926

TABLE II: Quantitative comparison with state-of-the-art methods. The best results are highlighted in bold . Methods prefixed with Ps- denote point-supervised approaches. mIoU and

P_{d}

are reported in (%), and

F_{a}

is reported in

10^{-5}

Fig. 5: Qualitative comparison of representative detection results on the SIRSTD-Pixel dataset. The GT panel shows the target location (top-right) and the corresponding ground-truth mask (bottom-right), displayed as a semi-transparent red overlay. Detection results from different methods are also visualized using semi-transparent red masks. Existing methods mainly suffer from three typical failure modes: missed detections under weak target responses, false alarms triggered by complex clutter, and inaccurate shape recovery with fragmented or biased masks. In the shown examples, the proposed method localizes targets more consistently and produces masks that more closely match the annotated thermal response region.

Seed Location	mIoU(%)	Area Ratio	Centroid Error	Radius Error
Center	87.19	1.1793	12.1526	0.2635
Random Interior	87.62	1.2432	14.6276	0.2673
Boundary	87.81	1.2530	15.0028	0.2654

TABLE III: Sensitivity analysis of PAMG under different seed locations on the SIRSTD-Pixel dataset.

(a) Impact of Fixed Support Radius $R_{s}$ (Blind Mode)
Radius $R_{s}$	2	4	8	12	16	20	24	28	32	40
mIoU (%)	8.75	34.43	72.69	85.09	88.37	88.34	87.73	87.03	86.25	84.99
(b) Impact of Adaptive Scale Factor $k$ (Guided Mode, $R_{s}=k\cdot r$ )
Scale Factor $k$	1	2	3	4	5	6	7	8	9	10
mIoU (%)	8.63	52.53	82.98	88.21	89.12	89.25	88.97	88.74	88.34	87.96

TABLE IV: Sensitivity analysis of the spatial support parameter in blind and guided PAMG modes on the SIRSTD-Pixel dataset. All values are reported in mIoU (%).

Fig. 6: Visual comparison of component-wise ablations in PAMG. Red overlays denote the generated pseudo masks. (a) Input image, where the top-right inset shows a magnified view of the target and the bottom-right inset shows a magnified clutter region. (b) Full PAMG accurately recovers the compact target while suppressing the clutter region. (c) Without the size prior, the target region is under-grown. (d) Without the saliency term, the target grows incompletely while clutter expands irregularly. (e) Without the homogeneity term, the target region expands into the background clutter. (f) Without the geometric prior, the clutter region grows without effective constraint.

Variant	mIoU (%)	Area Ratio
Full PAMG	88.34	1.22
w/o Size Prior	75.68	0.86
w/o Saliency	39.63	3.29
w/o Homogeneity	49.26	0.88
w/o Geometric Prior	75.02	13.40

TABLE V: Component-wise ablation study of PAMG on the SIRSTD-Pixel dataset. Area Ratio denotes the average ratio between the predicted mask area and the ground-truth area.

Fig. 7: Empirical validation of the theoretical generation boundary. Samples are grouped according to the boundary satisfaction ratio

\rho=\mathrm{SCR}/B(n,\gamma,R_{s})

. The average IoU and success rate (IoU

>0.5

) of the pseudo masks generated by PAMG are reported for each interval. As

\rho

increases, both metrics improve monotonically and gradually stabilize, indicating that targets satisfying the theoretical boundary condition lead to more reliable mask generation.

Model Variant	Key Change	Backbone Config		Detection Metrics			Shape
Model Variant	Key Change	Width	Max Stride	$P_{d}$ (%)	$F_{a}$ ( $10^{-5}$ )	AUC	mIoU (%)
Ablation on Backbone Capacity & Resolution
Variant-Width-S	Small Width	0.375	16	89.69	6.44	0.9485	34.63
Variant-Width-L	Large Width	0.75	16	90.49	12.00	0.9524	29.90
Variant-Stride	Standard Downsample	0.5	32	88.69	2.60	0.9435	39.27
Ablation on Temporal Modeling

TABLE VI: Ablation studies of RPR-Net on the SIRSTD-Pixel dataset. The “Key Change” column describes the modification relative to the baseline model.

Type	Method	Params $\downarrow$	FLOPs $\downarrow$	Latency $\downarrow$	AUC $\uparrow$
Single-frame	ACM [ 3 ]	0.40	2.01	2.95	0.8452
	AGPCNet [ 40 ]	12.36	215.91	37.75	0.8571
	ALCNet [ 4 ]	0.52	1.89	2.81	0.8775
	DNANet [ 17 ]	4.70	71.31	79.49	0.8390
	ISNet [ 39 ]	1.09	153.09	25.60	0.8710
	ISTDU-Net [ 10 ]	2.76	39.72	13.70	0.8577

TABLE VII: Efficiency comparison with state-of-the-art methods on the SIRSTD-Pixel dataset. Parameters, FLOPs, and latency are reported in M, G, and ms, respectively.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Abbreviation	Full Expression
IRSTD	Infrared Small Target Detection
PAMG	Physics-driven Adaptive Mask Generation
RPR-Net	Radius-aware Point Regression Network
PSF	Point Spread Function
MAP	Maximum A Posteriori
Hw-SNR	Homogeneity-Weighted Signal-to-Noise Ratio

Appendix

TABLE VIII: Main abbreviations used in this paper.

Infrared Small Target Detection

Infrared small target detection finds faint, often sub-pixel objects against cluttered thermal backgrounds while limiting false alarms.

See all 43 papers

Rethinking IRSTD: Single-Point Supervision Guided Encoder-only Framework is Enough for Infrared Small Target Detection
Apr 7, 2026Rixiang Ni, Boyang Li, Jun Chen +6Infrared Small Target Detection Small Object Detection
National University of Defense Technology, Changsha 410003, China
Exploring the Limits of End-to-End Feature-Affinity Propagation for Single-Point Supervised Infrared Small Target Detection
May 1, 2026Qiancheng Zhou, Wenhua ZhangInfrared Small Target Detection Object Detection
Learning with Semantic Priors: Stabilizing Point-Supervised Infrared Small Target Detection via Hierarchical Knowledge Distillation
May 14, 2026Yuanhang Yao, Ping Qian, Zhu Liu +2Infrared Small Target Detection Object Detection
School of Software Technology, Dalian University of Technology, China

Apr 7, 2026J/K move · Enter open · S save

Rethinking IRSTD: Single-Point Supervision Guided Encoder-only Framework is Enough for Infrared Small Target Detection

Rixiang Ni, Boyang Li, Jun Chen, Yonghao Li+5

National University of Defense Technology, Changsha 410003, China

Infrared Small Target Detection Small Object Detection Object Detection

Open paper AskPDF arXiv

Infrared small target detection (IRSTD) aims to separate small targets from clutter backgrounds. Extensive research is dedicated to the pixel-level supervision-guided "encoder-decoder" segmentation paradigm. Although having achieved promising performance, they neglect the fact that small targets only occupy a few pixels and are usually accompanied with blurred boundary caused by clutter backgrounds. Based on this observation, we argue that the first principle of IRSTD should be target localization instead of separating all target region accompanied with indistinguishable background noise. In this paper, we reformulate IRSTD as a centroid regression task and propose a novel Single-Point Supervision guided Infrared Probabilistic Response Encoding method (namely, SPIRE), which is indeed challenging due to the mismatch between reduced supervision network and equivalent output. Specifically, we first design a Point-Response Prior Supervision (PRPS), which transforms single-point annotations into probabilistic response map consistent with infrared point-target response characteristics, with a High-Resolution Probabilistic Encoder (HRPE) that enables encoder-only, end-to-end regression without decoder reconstruction. By preserving high-resolution features and increasing effective supervision density, SPIRE alleviates optimization instability under sparse target distributions. Finally, extensive experiments on various IRSTD benchmarks, including SIRST-UAVB and SIRST4 demonstrate that SPIRE achieves competitive target-level detection performance with consistently low false alarm rate (Fa) and significantly reduced computational cost. Code is publicly available at: https://github.com/NIRIXIANG/SPIRE-IRSTD.

Abstract

Infrared Small Target Detection

Explore similar work

Rethinking IRSTD: Single-Point Supervision Guided Encoder-only Framework is Enough for Infrared Small Target Detection

Exploring the Limits of End-to-End Feature-Affinity Propagation for Single-Point Supervised Infrared Small Target Detection

Learning with Semantic Priors: Stabilizing Point-Supervised Infrared Small Target Detection via Hierarchical Knowledge Distillation