Exploring the Limits of End-to-End Feature-Affinity Propagation for Single-Point Supervised Infrared Small Target Detection
| Authors | Qiancheng Zhou & Wenhua Zhang |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2605.00722 |
| Download | |
| Categories | cs.CV |
Abstract
Single-point supervised infrared small target detection (IRSTD) drastically reduces dense annotation costs. Current state-of-the-art (SOTA) methods achieve high precision by recovering mask supervision through explicit, offline pseudo-label construction, such as multi-stage active learning and physics-driven mask generation. In this paper, we study a minimalist alternative: generating point-to-mask supervision online through in-batch, point-anchored feature-affinity propagation. We instantiate this paradigm as GSACP, an end-to-end testbed that directly supervises the detector using hard-margin feature affinity gated by local image priors, entirely eliminating external label-evolution loops. This compact design, however, exposes an optimization bottleneck. Because the affinity target is generated from the same feature representation being optimized, training forms a self-referential loop. We theoretically formalize this as \emph{Self-Referential Propagation Drift}, a representation-supervision entanglement that can sharpen true boundaries or distort the feature space to satisfy its own targets. To systematically isolate these failure modes, we apply a protocolized single-variable ablation procedure spanning local EMA teacher decoupling, hard-background contrastive separation, and adaptive support geometry. On the SIRST3 dataset, GSACP-Final establishes a new ultra-low false-alarm operating regime, achieving a highly competitive mIoU while demonstrating a \mathrm{Fa}$) compared with PAL. By systematically deconstructing the end-to-end paradigm, we map its performance boundaries and show that in-batch feature propagation provides a compact alternative for deployment scenarios where false-alarm suppression is paramount.
Engineering Breakdown
Plain English
This paper addresses infrared small target detection (IRSTD) using only single-point labels per image, drastically reducing annotation costs compared to dense mask supervision. The authors propose GSACP, a method that generates mask supervision online during training by propagating feature affinity from annotated points using in-batch learning, eliminating the need for expensive offline pseudo-label construction pipelines like active learning or physics-driven mask generation. The key innovation is a hard-margin feature affinity mechanism gated by local image priors that directly supervises the detector end-to-end. While this approach achieves high precision with minimal annotation overhead, the authors identify an optimization bottleneck that the paper addresses through their compact design.
Core Technical Contribution
The core novelty is replacing offline, multi-stage pseudo-label construction with an online, point-anchored feature-affinity propagation mechanism that works entirely within a single training batch. Instead of evolving labels through external loops (active learning rounds or physics simulations), GSACP gates feature affinity using local image priors to convert single point annotations into mask-level supervision on-the-fly. This is fundamentally different from prior SOTA methods that rely on expensive label evolution—it's a minimalist, end-to-end approach that trades offline complexity for a tighter integration between annotation and learning. The hard-margin feature affinity design ensures that only pixels with sufficiently high feature similarity to the point annotation are supervised, providing a principled way to infer object boundaries from sparse labels.
How It Works
The pipeline begins with a single point annotation per infrared image (e.g., the center or a prominent pixel of a small target). During forward propagation, the detector extracts feature maps and computes pixel-wise feature affinity (similarity) relative to the annotated point's feature vector. The affinity scores are then gated by local image priors—low-level cues like intensity gradients or spatial smoothness constraints that reflect the structure of infrared imagery—to filter which pixels should receive supervision. Pixels exceeding a hard margin threshold (high affinity + high local prior) are labeled as target and used to compute the supervision signal; others are either negative or ignored. This gated propagation happens entirely in-batch and in-memory during training, eliminating the need to store pre-computed pseudo-masks or run external label-generation algorithms. The detector is trained end-to-end with this online mask supervision, learning to both detect targets and refine its own feature representation in tandem.
Production Impact
This approach has immediate practical value: annotation cost drops from dense pixel-level labeling to a single click per image, making it feasible to scale IRSTD systems to much larger datasets without prohibitive labeling budgets. Pipeline simplification is significant—removing offline pseudo-label stages (active learning loops, physics simulators, mask refinement tools) reduces engineering complexity and eliminates pipeline brittleness where one stage's errors cascade to the next. Latency impact is minimal because feature affinity computation and gating are O(HW) per image and added only during training; inference remains unchanged. The trade-off is that optimization becomes trickier (the paper identifies a bottleneck) and hyperparameter tuning for the hard margin and local prior weighting becomes critical to avoid supervising false positives. For teams building infrared surveillance or drone systems, this enables rapid model iteration on new domains where collecting dense annotations is impractical.
Limitations and When Not to Use This
The paper identifies but doesn't fully resolve an optimization bottleneck—the mechanism for balancing gated affinity targets with the detector's learned representation is underexplored, potentially making training unstable or slow to converge. The approach assumes that local image priors alone are sufficient to disambiguate targets from background, which may fail in cluttered infrared scenes where thermal noise or nearby distractors have similar feature statistics. The method is specialized to infrared imagery and leverages domain-specific priors (thermal signatures, local smoothness); generalization to other modalities (RGB, X-ray, SAR) is unclear and likely requires different gating mechanisms. Hard margin thresholding is brittle—setting the threshold too high misses target boundary pixels, too low introduces false positives—and no principled approach for tuning this per domain is presented. Evaluation on only infrared datasets limits understanding of robustness; failure modes on edge cases (occluded targets, extreme scale variation, multiple targets per image) are not thoroughly analyzed.
Research Context
This work builds on a growing body of research in weakly-supervised object detection (using image-level, point-level, or scribble-level labels instead of full masks) and sits at the intersection of semi-supervised learning and domain-specific computer vision. It directly improves upon prior IRSTD work that relied on active learning or physics-driven simulation to generate pseudo-masks, offering a lighter-weight alternative. The feature affinity idea echoes recent work in self-supervised and unsupervised segmentation (e.g., using feature similarity to propagate labels), but applies it in a new context with explicit hard-margin gating. This opens a research direction toward online, in-batch supervision mechanisms for other sparse annotation regimes, particularly in domains where annotation is expensive (medical imaging, satellite imagery, autonomous driving) and where domain priors can be readily incorporated.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::