Dynamic Local Geometry Capture in 3D Point Cloud Classification

Abstract

With the advent of PointNet, the popularity of deep neural networks has increased in point cloud analysis. PointNet’s successor, PointNet++, partitions the input point cloud and recursively applies PointNet to capture local geometry. PointNet++ model uses ball querying for local geometry capture in its set abstraction layers. Several models based on single scale grouping of PointNet++ continue to use ball querying with a fixed-radius ball. Due to its uniform scale in all directions, a ball lacks orientation and is ineffective in capturing complex local neighborhoods. Few recent models replace a fixed-sized ball with a fixed-sized ellipsoid or a fixed-sized cuboid to capture local neighborhoods. However, these methods are not still fully effective in capturing varying geometry proportions from different local neighborhoods on the object surface. We propose a novel technique of dynamically oriented and scaled ellipsoid based on unique local information to capture the local geometry better. We also propose ReducedPointNet++, a single set abstraction based single scale grouping model. Our model, along with dynamically oriented and scaled ellipsoid querying, achieves 92.1% classification accuracy on the ModelNet40 dataset. We achieve state-of-the-art 3D classification results on all six variants of the real-world ScanObjectNN dataset with an accuracy of 82.0% on the most challenging variant.