Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision (2304.01484v3)

Abstract: Training a convolutional neural network (CNN) to detect infrared small targets in a fully supervised manner has gained remarkable research interests in recent years, but is highly labor expensive since a large number of per-pixel annotations are required. To handle this problem, in this paper, we make the first attempt to achieve infrared small target detection with point-level supervision. Interestingly, during the training phase supervised by point labels, we discover that CNNs first learn to segment a cluster of pixels near the targets, and then gradually converge to predict groundtruth point labels. Motivated by this "mapping degeneration" phenomenon, we propose a label evolution framework named label evolution with single point supervision (LESPS) to progressively expand the point label by leveraging the intermediate predictions of CNNs. In this way, the network predictions can finally approximate the updated pseudo labels, and a pixel-level target mask can be obtained to train CNNs in an end-to-end manner. We conduct extensive experiments with insightful visualizations to validate the effectiveness of our method. Experimental results show that CNNs equipped with LESPS can well recover the target masks from corresponding point labels, {and can achieve over 70% and 95% of their fully supervised performance in terms of pixel-level intersection over union (IoU) and object-level probability of detection (Pd), respectively. Code is available at https://github.com/XinyiYing/LESPS.

• The paper introduces LESPS, a framework that evolves single-point annotations into full target masks through mapping degeneration.
• The method utilizes candidate pixel extraction, false alarm elimination, and weighted label updating to refine detection accuracy.
• Experiments demonstrate that LESPS achieves 70% to 95% of fully supervised performance, significantly reducing annotation labor.

Infrared Small Target Detection with Point Supervision

The paper "Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision" presents an innovative approach to infrared small target detection using point-level annotation rather than full pixel-level labeling. This research addresses a pivotal challenge in the field: the labor-intensive nature of achieving fully supervised training involving vast per-pixel annotations, particularly for infrared (IR) small target detection tasks.

Methodology

The authors introduce the concept of "mapping degeneration", where during training, Convolutional Neural Networks (CNNs) initially learn to segment clusters of pixels near the target with low confidence before refining to predict precise target locations. The authors utilize this observation to propose a novel framework known as Label Evolution with Single Point Supervision (LESPS). This framework effectively transforms point-supervised learning outcomes into approximate full-target masks over time by leveraging intermediate predictions from CNNs.

LESPS operates by dynamically expanding from single point annotations to wider mask labels during training iterations. The framework involves three key stages: candidate pixel extraction, false alarm elimination, and the weighted incorporation of predictions with current labels. With each label evolution and network training cycle, the method refines the pseudo labels, enabling the CNNs to move towards better delineation of target masks in an end-to-end manner.

Experimental Results

The experiments conducted within this paper demonstrate the effectiveness of the LESPS framework, with notable performance results. CNN models integrated with LESPS achieve between 70% to 95% of their fully supervised performance in terms of Intersection over Union (IoU) and probability of detection ($P_d$), respectively. Such results underscore the potential for significantly reducing the annotation workload while still maintaining a high level of accuracy in detecting small infrared targets.

Implications and Future Directions

The positive outcomes from this research indicate that LESPS could offer a much-needed solution to the prevalent bottleneck of annotation in infrared target detection tasks. This advancement has marked implications for civil and military applications, including but not limited to traffic monitoring, maritime rescue, and military surveillance.

Looking forward, the application of LESPS could be explored in broader contexts beyond just infrared small target detection. For instance, similar label evolution strategies could potentially benefit other domains where full annotations are impractical due to cost or complexity, particularly in fields like satellite imagery analysis and biomedical image segmentation.

The theoretical implications of this work also point towards deeper explorations of the relationship between weak supervision and model robustness. Understanding how neural networks evolve their learning during training phases with minimal supervision could further enhance model efficiency and effectiveness in various domains, feasibly influencing the development of more scalable AI systems.

In conclusion, the paper's approach of mapping degeneration and label evolution opens promising avenues for both theoretical exploration and practical application in infrared imaging and beyond. The code and additional resources being made publicly available will undoubtedly facilitate continued research and development in this exciting area.