TreeGCN-ED: a tree-structured graph-based autoencoder framework for point cloud processing

Abstract

Point cloud is a widely used technique for representing and storing 3D geometric data. Several methods have been proposed for processing point clouds for tasks such as 3D shape classification and clustering. This work presents a tree-structured autoencoder framework to generate robust embeddings of point clouds through hierarchical information aggregation using graph convolution. We visualize the t-SNE map to highlight the ability of learned embeddings to distinguish between different object classes. We further demonstrate the robustness of these embeddings in applications such as point cloud interpolation, completion, and single image-based point cloud reconstruction. The anonymized code is available here for research purposes.