Hierarchical Object Relationship Constrained Monocular Depth Estimation.

Monocular depth estimation has been gaining growing momentum in recent years. Despite significant advances of this task, due to the inherent difficulty reliably capturing contextual cues from RGB images, it remains challenging accurately predict scenes with complicated and cluttered spatial arrangement objects. Instead naively utilizing primary features single image, paper we propose a hierarchical object relationship constrained network for monocular estimation, which could enable accurate smooth prediction image. The key idea our method is exploit object-centric as constraints compensate regularity changing. In particular, design semantics-guided CNN encode original image into global context feature map objects’ local simultaneously, so that can leverage such effective consolidated coding scheme over scenario samples guide more way. Benefiting local-to-global constraints, well respect changing preserve details at same time. addition, approach make full use semantic across inner-object components neighboring objects define constraints. We conduct extensive experiments comprehensive evaluations on widely-used public datasets, confirm outperforms most state-of-the-art methods preserving depth.