Weakly Supervised Image Segmentation

Weakly supervised image segmentation studies how to produce dense pixel masks when detailed pixel-by-pixel annotations are scarce, costly, or unreliable. Common approaches turn image labels into class activation maps and pseudo-labels, propagate sparse annotations across regions, or use geometric, anatomical, and language-derived constraints; foundation models and vision-language features increasingly provide extra guidance. The methods are tested on standard semantic-segmentation benchmarks as well as medical, industrial, and remote-sensing images, where label types and structures vary widely. A central challenge is recovering complete objects and precise boundaries from incomplete or noisy cues while maintaining reliable performance across domains and annotation settings.