Deep Learning 3D Shapes Using Alt-az Anisotropic 2-Sphere Convolution

Authors: Min Liu, Fupin Yao, Chiho Choi, Ayan Sinha, Karthik Ramani

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate theoretically and experimentally that our proposed method has the possibility to bridge the gap between 2D images and 3D shapes with the desired rotation equivariance/invariance, and its effectiveness is evaluated in applications of non-rigid/ rigid shape classification and shape retrieval.
Researcher Affiliation Collaboration Min Liu Purdue University Fupin Yao Purdue University Chiho Choi Honda Research Institute, USA. Ayan Sinha Magic Leap Inc. Karthik Ramani Purdue University
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access information (e.g., specific link, explicit statement of code release) for open-source code for the methodology described.
Open Datasets Yes We first conduct experiments on SHREC 11 non-rigid shape classification. We further experiment on Model Net10 and Model Net40 rigid shape databases. Finally, we run shape retrieval experiments on Shape Net Core55, following rules of the SHREC 17 3D shape contest (Savva et al. (2016).)
Dataset Splits No In each category, 16 objects are used for training and 4 objects are used for testing. The paper mentions training and testing sets but does not explicitly provide details about validation splits (e.g., percentages, counts, or methodology for a dedicated validation set).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, cloud instances) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers).
Experiment Setup Yes Our network contains five a3SConv-dropout-Re LU-LMP blocks. A 20% dropout is added right after each spherical convolution layer for regularization. The resulting spherical functions are pooled using a global max pooling (GMP) layer followed by two fully connected layers for the final classification. A 50% dropout layer is inserted in between the last two fully connected layers. We use 32, 64, 64, 128, 128 features for the a3SConv layers, and 512 features are output from the GMP and fed into the first fully connected layer. Each filter on S2 has kernel size ring-2, stride 1 and each LMP layer has size ring-2 and stride 2.