2.5D visual relationship detection

Visual 2.5D perception involves understanding the semantics and geometry of a scene through reasoning about object relationships with respect to viewer. However, existing works in visual recognition primarily focus on semantics. To bridge this gap, we study relationship detection (2.5VRD), which goal is jointly detect objects predict their relative depth occlusion relationships. Unlike general VRD, 2.5VRD egocentric, using camera’s viewpoint as common reference for all estimation, object-centric does not only depth. enable progress task, construct new dataset consisting 220K human-annotated among 512K from 11K images. We analyze conduct extensive experiments including benchmarking multiple state-of-the-art VRD models task. Experimental results show that largely rely semantic cues simple heuristics solve 2.5VRD, motivating further research perception. will make our source code publicly available.