Occlusion Reasoning

Occlusion reasoning studies how vision systems infer what is hidden, which objects are visible, and how visibility changes across views or time, supporting reliable perception and action in cluttered scenes. Approaches include multi-view geometric consistency, depth and Z-order modeling, amodal shape completion, scene graphs, and temporal tracking. Evaluation spans stereo benchmarks such as KITTI, tracking datasets such as MOT, occlusion-aware segmentation benchmarks, and autonomous-driving settings such as Waymo. Task-specific methods can improve perception and planning, but hidden-state ambiguity remains difficult: vision-language models often miss direct physical cues, and detector confidence may indicate object presence without showing whether an instance is visible.