Designing efficient deep learning models for 3D point cloud perception is becoming a major research direction. Point-voxel convolution (PVConv) Liu et al. (2019) is a pioneering research work in this topic. However, since with quite a few layers of simple 3D convolutions and linear point-voxel feature fusion operations, it still has considerable room for improvement in performance. In this paper, we propose a novel pyramid point-voxel convolution (PyraPVConv) block with two key structural modifications to address the above issues. First, PyraPVConv uses a voxel pyramid module to fully extract voxel features in the manner of feature pyramid, such that sufficient voxel features can be obtained efficiently. Second, a sharable attention module is utilized to capture compatible features between multi-scale voxels in pyramid and point cloud for aggregation, as well as to reduce the complexity via structure sharing. Extensive results on three point cloud perception tasks, i.e., indoor scene segmentation, object part segmentation and 3D object detection, validate that the networks constructed by stacking PyraPVConv blocks are efficient in terms of both GPU memory consumption and computational complexity, and are superior to the state-of-the-art methods.
Copyright © 2022 Yuhong Chen et al.