SPSIS: Single-Point Supervised Instance Segmentation for Remote Sensing

Remote sensing instance segmentation is a significant but difficult task due to the need for a number of accurate mask labels. Although existing weakly supervised methods use multiple points or bounding boxes for one object to reduce labeling requirements, they still require many manual labeling costs. Therefore, we propose an SPSIS model with only one point for each object to reduce the annotation burden. Pseudo bounding boxes are firstly generated using the refined masks of the Segment Anything Model and candidate points are randomly sampled within the boxes. In addition, we propose a point classification method combining ensemble learning and label propagation algorithm to classify sampled points. Finally, we use a point loss function so that the mask-based instance segmentation model can effectively adapt to the point samples. We have conducted extensive experiments on AG and WHU datasets, demonstrating the superiority of SPSIS. In addition, SPSIS significantly lessens the precision difference between weakly and fully supervised instance segmentation.