Abstract
Designing efficient deep learning models for 3D point clouds is an important research topic. Point-voxel convolution (Liu et al. in NeurIPS, 2019) is a pioneering approach in this direction, but it still has considerable room for improvement in terms of performance, since it has quite a few layers of simple 3D convolutions and linear point-voxel feature fusion operations. To resolve these issues, we propose a novel reparameterizable point-voxel convolution (RepPVConv) block. First, RepPVConv adopts two reparameterizable 3D convolution modules to extract more informative voxel features without introducing any extra computational overhead for inference. The rationale is that the reparameterizable 3D convolution modules are trained in high-capacity modes but are reparameterized into low-capacity modes during inference while losslessly maintaining the original performance. Second, RepPVConv attentively fuses the reparameterized voxel features with those of points. Since the proposed approach operates in a nonlinear manner, descriptive reparameterized voxel features can be better utilized. Extensive experimental results show that RepPVConv-based networks are efficient in terms of both GPU memory consumption and computational complexity and significantly outperform the state-of-the-art methods.