Hard Example Mining-based Training Strategy for Pointly-Supervised Instance Segmentation

The time-consuming task of generating mask-level annotations has impeded the extensive adoption of instance segmentation. To address this challenge, weakly-supervised instance segmentation has emerged, leveraging more available annotation types to achieve performance comparable to fully-supervised approaches. Recent advancements in pointly-supervised instance segmentation demonstrate promising outcomes by employing a mere 20 labeled points within bounding box per object. This approach substantially decreases the annotation burden while preserving competitive performance levels. However, its indiscriminate training strategy for all labeled points severely limits the segmentation performance of the model. In this paper, we propose a novel Hard Example Mining-based Training Strategy for pointly-supervised instance segmentation, effectively utilizing limited point-based supervision and focusing on challenging labeled points. Our method addresses the need for efficient supervision by leveraging sparse point annotations while maintaining competitive performance. On COCO, we show that models trained with our strategy yield significantly gain compared against normal training strategy.