Reasoning Segmentation

Reasoning segmentation asks vision-language systems to identify the pixels belonging to targets described by complex instructions, scene context, or multi-step intent. Common approaches combine multimodal language models with chain-of-thought reasoning, attention or prompt-based localization, candidate-mask selection, and decoders such as SAM; some separate semantic interpretation from spatial grounding. Training uses supervised examples and increasingly reinforcement learning, while benchmarks cover still images and videos as well as settings such as medical, remote-sensing, and adverse-weather imagery. A central tension is how to improve reliable spatial and temporal grounding while keeping reasoning interpretable, efficient, and robust to unfamiliar queries and conditions.