SubConvFusion3D: A Multi-Scale Sub-Convolutional Networks and Feature Fusion for 3D Deep Learning

Convolutional neural networks (CNNs) are useful tools for analyzing images in several facets of computer vision. Compared to 2D images, 3D points are not represented as a regular grid but rather as a collection of irregularly spaced points in three dimensions, varying significantly depending on the object or scene being described. This makes applying convolutions to 3D points challenging. This work presents a new method called SubConvFusion3D for applying convolutions on a 3D point cloud. By considering the multiple scales of the 3D points, the proposed method minimizes confusion and redundancy while extracting different features at different scales and detail levels. To achieve this, different sub-convolutional networks are first used, allowing for independent analysis. It enables 1D convolutions to extract features for every subnetwork at various scales. The provided features are finally combined to form the final feature vector, by using a new merging (fusion) function. The experimental results conducted on the ModelNet40 dataset show competitive performance achieved by our methodology. The training of SubConvFusion3D achieved an outstanding training accuracy of 87.2%, demonstrating our approach's competitive performance.