Infrared small-target detection (IRSTD) has benefited from deep learning, yet most existing methods still rely on dense pixel-level annotations, which are costly and often unreliable for tiny and weak targets. Point supervision offers a more practical alternative, but current point-supervised methods usually construct pseudo-labels based on the distance between pixels and annotated points or cluster centers, which introduces spatial bias and may miss genuine target pixels away from these reference points. To address this issue, we propose GMVG-DIR, a point-supervised IRSTD framework that combines Gradient-Guided Minimum Variance Growth (GMVG) with Deep Iterative Refinement (DIR). GMVG first estimates target likelihood from gradient-guided aggregation of contour closure and edge responses and then converts it into structurally coherent pseudo-labels via the Minimum Variance Growth filter, without relying on distance cues. DIR further improves the pseudo-labels by incorporating reliable semantic guidance into an iterative refinement process, thereby reducing error propagation. By emphasizing structural consistency rather than spatial proximity, the proposed framework better preserves irregular target shapes and remains robust to point-label deviation. Extensive experiments on NUDT-SIRST, IRSTD-1k, and NUAA-SIRST show that GMVG-DIR improves pseudo-label fidelity and achieves competitive point-supervised performance across multiple dataset-backbone settings, especially in IoU and Pd.
Shi et al. (Thu,) studied this question.