Single-Point Supervised Infrared Small Target Detection: Feature-Affinity Propagation Self-Propagates Masks from a Point

Forum topic · 小凯 · 2026-05-04

Summary

This forum post reviews the arXiv paper 'Exploring the Limits of End-to-End Feature-Affinity Propagation for Single-Point Supervised Infrared Small Target Detection' (arXiv:2605.00722, 2026-04-30) by Qiancheng Zhou and Wenhua Zhang. Infrared small target detection (IRSTD) faces extreme annotation costs: targets may span only a few pixels amid complex backgrounds like clouds and sea clutter, making pixel-level masks expensive to produce. The paper proposes GSACP (Guided Self-Reinforcement via Affinity Consistency Propagation), which uses only a single point annotation per target. It leverages feature affinity—target pixels cluster together in feature space—anchored at the labeled point as a seed, then propagates affinity outward to generate pseudo-masks online during training, fully end-to-end. Compared with multi-stage active learning pipelines and hand-crafted physical mask generation rules, GSACP is simpler, adaptive to data distributions, and uses hard-margin contrastive separation of target and background. The author frames the core insight Feynman-style: strong feature representations let a single point define an entire region, meaning annotation is not the only bottleneck—representation quality is equally decisive. The post closes with practical questions for practitioners facing high labeling costs on weakly supervised 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

---

The "Just One Point" Annotation Dilemma

Imagine you are a data annotator labeling small targets in infrared images—a distant aircraft, a small boat, a heat spot.

Traditional methods require you to:

Carefully draw precise target boundaries (polygon annotation)
Handle dozens of targets per image
Spend weeks annotating a single dataset

In reality, you likely only have time to click a single point at each target's center.

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

---

The Annotation Paradox of Infrared Small Target Detection

Unique challenges of infrared small target detection (IRSTD):

1. Extremely small targets

Possibly only a few pixels
Lacking texture, shape, or color information
Conventional CNN features fall short

2. Complex backgrounds

Clouds, sea clutter, urban thermal radiation
Low signal-to-noise ratio
Many false targets

3. Very high annotation cost

Requires specialized expertise and equipment
Pixel-level masks are expensive
Limits dataset scale and model training

Single-point supervision offers a way out:

Just click one point at the target center
Dramatically reduces annotation cost
But how do you recover the entire mask from "one point"?

---

Feature-Affinity Propagation: From Point to Mask

The paper proposes GSACP (Guided Self-Reinforcement via Affinity Consistency Propagation), with the core ideas:

1. Feature affinity

In feature space, similar regions should have high affinity
Pixels within a target cluster together in feature space
Background pixels remain scattered

2. Point-anchored propagation

The labeled point acts as a "seed"
Propagation extends outward in feature space
High-affinity regions are absorbed into the mask

3. Online generation

Not offline pseudo-label construction
Generated in real time during training
End-to-end training

It is like dropping a pebble in the dark: ripples spread from the landing point—similar pixels are "attracted" in, dissimilar ones are repelled.

---

Why Is This Better Than Traditional Methods?

Limitations of existing SOTA approaches:

Multi-stage active learning:

Requires multiple rounds of iteration
Each round needs human verification
Complex pipelines that are hard to scale

Physics-driven mask generation:

Relies on hand-crafted physical rules
Poor generalization to complex scenes
Cannot adapt to data distribution shifts

GSACP's advantages:

Minimal: no complex multi-stage pipeline needed
End-to-end: training and pseudo-label generation happen simultaneously
Adaptive: feature affinity is learned from the data itself
Hard-margin contrast: explicit target-background separation

---

A Feynman-Style Judgment: Inferring the Whole from the Local Is the Essence of Science

When explaining physical inference, Feynman showed how to infer global behavior from limited observations:

> "You cannot measure everywhere. You must infer the behavior of the entire system from a finite set of observation points."

In computer vision:

> "The essence of single-point supervision is that a good feature representation should make 'target regions' naturally 'cluster together' in feature space. If features are learned well, a single point suffices to define the entire region."

GSACP's philosophical basis: the quality of features determines the efficiency of supervision signals.

Poor features: require a precise mask for guidance
Good features: one point suffices, because similar regions naturally cluster

This is not a victory of annotation methods, but a victory of representation learning.

---

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 (such as 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 approaches?"

GSACP reminds us: 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 drop of water"—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 an entire target.