AVS-Net: Point Sampling with Adaptive Voxel Size for 3D Scene Understanding

The recent advancements in point cloud learning have enabled intelligentvehicles and robots to comprehend 3D environments better. However, processinglarge-scale 3D scenes remains a challenging problem, such that efficientdownsampling methods play a crucial role in point cloud learning. Existingdownsampling methods either require a huge computational burden or sacrificefine-grained geometric information. For such purpose, this paper presents anadvanced sampler that achieves both high accuracy and efficiency. The proposedmethod utilizes voxel centroid sampling as a foundation but effectivelyaddresses the challenges regarding voxel size determination and thepreservation of critical geometric cues. Specifically, we propose a VoxelAdaptation Module that adaptively adjusts voxel sizes with the reference ofpoint-based downsampling ratio. This ensures that the sampling results exhibita favorable distribution for comprehending various 3D objects or scenes.Meanwhile, we introduce a network compatible with arbitrary voxel sizes forsampling and feature extraction while maintaining high efficiency. The proposedapproach is demonstrated with 3D object detection and 3D semantic segmentation.Compared to existing state-of-the-art methods, our approach achieves betteraccuracy on outdoor and indoor large-scale datasets, e.g. Waymo and ScanNet,with promising efficiency.