H-CNN: Spatial Hashing Based CNN for 3D Shape Analysis

Abstract

We present a novel spatial hashing based data structure to facilitate 3D shape analysis using convolutional neural networks (CNNs). Our method well utilizes the sparse occupancy of 3D shape boundary and builds hierarchical hash tables for an input model under different resolutions. Based on this data structure, we design two efficient GPU algorithms namely hash2col and col2hash so that the CNN operations like convolution and pooling can be efficiently parallelized. The spatial hashing is nearly minimal, and our data structure is almost of the same size as the raw input. Compared with state-of-the-art octree-based methods, our data structure significantly reduces the memory footprint during the CNN training. As the input geometry features are more compactly packed, CNN operations also run faster with our data structure. The experiment shows that, under the same network structure, our method yields comparable or better benchmarks compared to the state-of-the-art while it has only one-third memory consumption. Such superior memory performance allows the CNN to handle high-resolution shape analysis.

Dual Octree Graph Networks for Learning Adaptive Volumetric Shape Representations

“…Sparse-Voxel-based CNNs. The drawback of volumetric CNNs is overcome by sparse-voxel-based CNNs, which adopt spatially adaptive data structures such as octrees [Wang et al 2017] and hash tables [Choy et al 2019;Graham et al 2018;Shao et al 2018] to index non-empty voxels efficiently and constrain the convolution over these sparse voxels. A set of sparse-voxel-based CNNs have been proposed for shape reconstruction and generation, where an encoder-decoder network is learned to map input point cloud to octrees with occupancy values [Häne et al 2017;Tatarchenko et al 2017], adaptive planar patches [Wang et al 2018b], or moving-least-squares points .…”

“…Although these approaches can easily adapt deep neural networks developed for 2D images to 3D learning, their memory and computational costs grow cubically as the volumetric resolution increases, making them difficult to model 3D shape details. A set of methods [Choy et al 2019;Graham et al 2018;Shao et al 2018;Wang et al 2017] represent 3D shapes with sparse non-empty voxels and design neural networks that operate only on sparse voxels. Although these sparse-voxel-based methods significantly reduce computational and memory cost, the features in empty voxels are ignored or simply set to zero, and predicting the locations of sparse voxels for shape generation and reconstruction is a difficult task, especially for incomplete inputs.…”

Two Heads are Better than One: Geometric-Latent Attention for Point Cloud Classification and Segmentation

“…Methods like OctNet [22] and O-CNN [31] save computation time by using octrees to avoid processing empty spaces. [10] and [23] use Kd-tree and Hash structures instead. [26] uses sparse 3D convolutions rather than efficient data structures.…”

“…Some approaches project the 3D raw data into a regular structure (e.g. voxels) where 3D convolutions can be used [10,17,22,23,31,39]. Other approaches use multilayer perceptrons (MLP) to process point clouds directly [19,20,29].…”

Recognition of Point Sets Objects in Realistic Scenes

“…Methods based on the 3D voxel grid data are given as follows: this method works by meshing or voxelizing various 3D data and then designing the corresponding 3D convolutional neural network for feature extraction and recognition. References [1,[4][5][6][7] is a series of convolutional neural network algorithms whose input data is a voxel grid, but these algorithms all consume a lot of computational costs because of the sparseness of the data and the features of convolution in 3D. e requirements for resolution are high.…”