Uncertainty-Aware Weakly Supervised Building Change Detection with Point Annotations

Abstract

Point-based weakly supervised strategies can reduce annotation costs but suffer from the problems of semantic sparsity and boundary ambiguity. This paper proposes an uncertainty-aware weakly supervised building change detection method with point annotations. First, the method leverages the proposal generation capability and embedding space of the Segment Anything model to generate reliable three-class pseudo-labels. By fusing appearance similarity and geometric overlap at the proposal level, it distinguishes high-confidence changed/no-change regions and uncertain regions, which not only enriches the semantic information of annotations but also mitigates the prompt-induced semantic bias. Then, an entropy minimization strategy is adopted to impose constraints on low-confidence uncertain pixels, which suppresses noisy supervision while achieving refined boundary extraction. To verify the effectiveness of the proposed method, experiments on standard building change detection benchmark data sets show that compared with existing weakly supervised methods, the proposal-level matching strategy significantly improves the fidelity of pseudo-labels, thus yielding more robust and accurate change detection results.