Referring Video Object Segmentation

Referring video object segmentation (RVOS) studies how to turn a text, speech, or other prompt into pixel-level masks for the intended object across video frames. Common approaches combine vision-language models or multimodal large language models for query understanding with segmentation and tracking systems such as SAM, using object memories, key frames, motion cues, and mask propagation to maintain identity over time. Systems are trained on datasets such as Ref-YouTube-VOS, MeViS, DAVIS, and ReVOS, while training-free pipelines increasingly use object tracks and iterative reasoning; newer evaluations also test audio queries, absent targets, long untrimmed videos, and causal online settings. A central challenge is balancing accurate language grounding and stable masks under motion, occlusion, ambiguity, and target reappearance against the extra computation required for repeated reasoning or refinement.