MonoGRNet: A Geometric Reasoning Network for Monocular 3D Object Localization

• DOI: 10.1609/aaai.v33i01.33018851
• Publication Date: 2019-08-20T07:34:37Z
• Authors: Zengyi Qin; Jinglu Wang; Yan Lu
• Keywords: Minimum bounding box; Amodal perception; Bounding overwatch; Monocular; RGB color model

Localizing objects in the real 3D space, which plays a crucial role scene understanding, is particularly challenging given only single RGB image due to geometric information loss during imagery projection. We propose MonoGRNet for amodal object localization from monocular via reasoning both observed 2D projection and unobserved depth dimension. single, unified network composed of four task-specific subnetworks, responsible detection, instance estimation (IDE), local corner regression. Unlike pixel-level that needs per-pixel annotations, we novel IDE method directly predicts targeting bounding box’s center using sparse supervision. The further achieved by estimating position horizontal vertical dimensions. Finally, jointly learned optimizing locations poses boxes global context. demonstrate achieves state-of-the-art performance on datasets.