Weakly Supervised Semantic Segmentation

Weakly supervised semantic segmentation studies how to assign a class to every pixel when training data lacks complete pixel-level masks, reducing the cost of annotation. Common approaches turn image labels into class activation maps, refine noisy pseudo-masks with self-training or consistency losses, and use boundaries, spatial priors, or vision-language foundation models to improve localization. Methods are tested on benchmarks such as PASCAL VOC and MS COCO, as well as medical, satellite, and other domain-specific imagery with image-level labels, points, or scribbles. A central challenge is that sparse cues often highlight only the most distinctive parts of objects, so better coverage and boundaries must be balanced against noisy pseudo-labels and limited generalization across domains.