SI-RCNN: A Shape-Invariant Set-Abstraction for 3D Object detection
Xiaoping Wu, Guiyang Luo, Quan Sheng Yuan, Jinglin Li, Fangchun Yang · 2021
Recent years have witnessed more attention being paid to 3D object detection, due to its indispensable roles in autonomous driving and intelligent transportation systems. The accuracy of 3D object detection heavily depends on sufficient and robust feature extractors. However, most feature extractors adopt fixed perceptive fields or fixed spatial locations, which could not efficiently model geometric variations or geometric transformations in object scale, pose, and part deformation. To fill this gap, this paper proposes a novel Shape-Invariant Set-Abstraction network (SI-RCNN) for 3D object detection in point clouds, which can extract shape-adaptive features. Specifically, we first propose an Offsets Generation Module to learn a set of offsets, where the shape of the offsets can adapt to the changes of features. Then, trilinear interpolation is applied to assign features to each offset, since it might be empty due to discrete points in the cloud data. Finally, we further exploit the ground-truth box to supervise the offsets. Extensive experiments demonstrate the effectiveness of our proposed SI-RCNN on the competitive KITTI 3D detection benchmark, as compared to several state-of-the-art methods.