OPA-3D: Occlusion-Aware Pixel-Wise Aggregation for Monocular 3D Object Detection

Monocular 3D object detection has recently made a significant leap forward thanks to the use of pre-trained depth estimators for pseudo-LiDAR recovery. Yet, such two-stage methods typically suffer from overfitting and are incapable explicitly encapsulating geometric relation between bounding box. To overcome this limitation, we instead propose jointly estimate dense scene with depth-bounding box residuals boxes, allowing two-stream objects that harnesses both geometry context information. Thereby, stream combines visible recover via explicit occlusion-aware optimization. In addition, based projection scheme is employed in an effort enhance distance perception. The second stream, named as Context Stream, directly regresses location size. This novel representation enables us enforce cross-stream consistency terms, which aligns outputs streams, further improves overall performance. Extensive experiments on public benchmark demonstrate OPA-3D outperforms state-of-the-art main Car category, whilst keeping real-time inference speed.