Fully supervised instance segmentation typically demands large datasets with pixel-wise annotations, which are costly and slow to produce. Weakly and semi-supervised alternatives use cheaper labels but sacrifice accuracy. This paper proposes a hybrid pipeline that combines economical bounding-box labels with a smaller set of accurate pixel-wise labels, aiming to save annotation effort without a major performance drop.
The core idea is to learn label noise suppression. Two ancillary models are trained: one suppresses mislabeling between foreground and background, and the other suppresses noise from mislabeling individual instances. These models help generate accurate pseudo pixel-wise labels from bounding boxes. Additional components—category-aware spatial attention, category and instance constraint modules, and a self-learning scheme—further refine the pseudo labels.
Experiments on PASCAL VOC 2012 and Cityscapes show the method achieves competitive results compared with fully supervised approaches while requiring much less labeling cost. The paper, published in Neurocomputing, is available on arXiv as 2610.20412.