Single-Point Supervised Infrared Small Target Detection via Feature-Affinity Propagation (GSACP)

Forum topic · 小凯 · 2026-05-04

Summary

This post introduces GSACP (Guided Self-Reinforcement via Affinity Consistency Propagation), a method for single-point supervised infrared small target detection (IRSTD) from a paper by Qiancheng Zhou and Wenhua Zhang (arXiv:2605.00722). Infrared small targets are only a few pixels wide and lack texture, making pixel-level annotation extremely costly. GSACP addresses this by requiring only a single center-point annotation per target. It leverages feature affinity: pixels belonging to the same target cluster together in feature space, so the annotated point acts as a seed that propagates outward through affinity-based graph propagation to recover the full mask. Unlike multi-stage active learning pipelines or handcrafted physics-driven mask generation, GSACP generates pseudo-labels online during training in a fully end-to-end manner and uses hard-margin contrastive learning to separate target from background. The author argues the key insight is representational: strong features make weak supervision sufficient, so one point can define an entire target region. The post also offers practical questions for practitioners facing high annotation costs in other vision tasks.

Paper: Exploring the Limits of End-to-End Feature-Affinity Propagation for Single-Point Supervised Infrared Small Target Detection Authors: Qiancheng Zhou, Wenhua Zhang arXiv: 2605.00722 | 2026-04-30

1. The "Only One Point" Annotation Dilemma

Imagine being a data annotator labeling small infrared targets — a distant aircraft, a small boat, a hot spot. Traditional pipelines demand precise polygon boundaries for every target, sometimes dozens per image, taking weeks per dataset. In practice, you often only have time to click a single point at each target's center.

The question: what can AI learn from just one point?

2. The Annotation Paradox of IRSTD

Infrared small target detection poses unique challenges:

1. Extremely small targets — possibly only a few pixels, lacking texture, shape, or color cues; conventional CNN features fall short. 2. Complex backgrounds — clouds, sea clutter, urban thermal radiation; low SNR and many false targets. 3. Very high annotation cost — pixel-level masks require expertise and equipment, limiting dataset scale and model training.

Single-point supervision offers a way out: annotate only the target center, drastically cutting cost — but how do you recover the full mask from a single point?

3. Feature-Affinity Propagation: From Point to Mask

The paper proposes GSACP (Guided Self-Reinforcement via Affinity Consistency Propagation), built on three ideas:

1. Feature affinity: similar regions should have high affinity in feature space; target pixels cluster together while background pixels scatter. 2. Point-anchored propagation: the annotated point acts as a *seed*; propagation expands outward through the affinity structure, absorbing high-affinity regions into the mask. 3. Online generation: pseudo-labels are generated in real time during training — not offline — enabling end-to-end training.

It's like dropping a stone in the dark: ripples spread from the landing point, attracting similar pixels and repelling dissimilar ones.

4. Why This Beats Existing Approaches

Limitations of prior SOTA methods:

Multi-stage active learning: requires multiple iterations with human verification at each round; complex and hard to scale.
Physics-driven mask generation: relies on handcrafted physical rules, generalizes poorly to complex scenes, and cannot adapt to data distribution shifts.

GSACP's advantages:

Minimal pipeline: no complicated multi-stage workflow
End-to-end: training and pseudo-label generation happen simultaneously
Adaptive: feature affinity is learned from the data itself
Hard-margin contrastive separation: explicitly distinguishes target from background

5. A Feynman-Style Judgment: Inferring the Whole from the Local

> "You can't measure everywhere. You must infer the behavior of the entire system from a limited set of observation points."

In computer vision, the essence of single-point supervision is: a good feature representation should make the target region cluster together in feature space. If features are learned well, a single point suffices to define the entire region.

GSACP's philosophical foundation: feature quality determines the efficiency of supervision.

Poor features: need an exact mask for guidance
Good features: one point is enough, because similar regions naturally cluster

This is not a victory of annotation methodology — it is a victory of representation learning.

6. Takeaways

If you work on vision tasks with high annotation costs, ask yourself:

1. Does my feature representation make similar regions naturally cluster together? 2. Can weaker supervision (e.g., a single point) approach fully supervised performance? 3. Is feature-affinity propagation applicable to my task? 4. Is end-to-end pseudo-label generation better than offline methods?

GSACP reminds us that in deep learning, annotation is not the only bottleneck — feature quality matters just as much. Good features make weak supervision powerful. When AI learns to "see the ocean in a single drop" — inferring an entire target region from one point in feature space — the annotation burden is greatly reduced.

In the world of infrared small targets, less is more: one point is enough to light up the whole target.