Weakly-supervised remote sensing salient object detection using point supervision

Abstract

The progress of salient object detection in optical remote sensing images largely depends on full pixel-level annotations, which require a time-consuming and labor-intensive labeling process. In contrast, point annotations are particularly appealing because they require minimal manual effort. In this work, we investigate point-supervised remote sensing salient object detection by leveraging point-activated pseudo-mask generation and adapted network design. Segment Anything Model (SAM) can generate object masks from point annotations, enabling the direct use of these masks as supervisory signals for network training. However, our empirical studies show that naively applying SAM often leads to masks that incorrectly include both objects and their cast shadows as foreground in remote sensing scenarios. To address this issue, we propose a simple mean-shift thresholding method that removes shadow regions based on the statistical histogram. This strategy substantially improves the quality of the generated pseudo masks. Meanwhile, we introduce a Gated Local Block Attention (GLBA) module, which can be effectively supervised by the generated pseudo-mask to suppress background noise and enhance salient features in complex remote sensing scenes. Integrated with this module, our architecture shows promising results in salient object detection. Experimental evaluations on three remote sensing benchmarks demonstrate the effectiveness of the proposed point-supervised method, which performs favorably against the existing fully supervised and scribble-supervised methods.