Learning Representations and Generative Models for 3D Point Clouds

Authors: Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, Leonidas Guibas

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we perform comprehensive experiments evaluating all of our models both quantitatively and qualitatively.
Researcher Affiliation Academia 1Department of Computer Science, Stanford University, USA 2MILA, Department of Computer Science and Operations Research, University of Montr eal, Canada.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Last, the code for all our models is publicly available1. 1http://github.com/optas/latent_3d_points
Open Datasets Yes In all experiments in the main paper, we use shapes from the Shape Net repository (Chang et al., 2015), that are axis aligned and centered into the unit sphere.
Dataset Splits Yes Unless otherwise stated, we train models with point clouds from a single object class and work with train/validation/test sets of an 85%-5%-10% split.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or cloud computing resources used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details, such as library names with version numbers, needed to replicate the experiment.
Experiment Setup Yes The input to our AE network is a point cloud with 2048 points (2048 3 matrix). The encoder architecture follows the design principle of (Qi et al., 2016a): 1-D convolutional layers with kernel size 1 and an increasing number of features; this approach encodes every point independently. In our implementation we use 5 1-D convolutional layers, each followed by a Re LU (Nair & Hinton, 2010) and a batch-normalization layer (Ioffe & Szegedy, 2015). Our decoder transforms the latent vector using 3 fully connected layers, the first two having Re LUs, to produce a 2048 3 output. In the remainder of the paper, unless otherwise stated, we use an AE with a 128-dimensional bottleneck layer. The architecture of the discriminator is identical to the AE (modulo the filter-sizes and number of neurons), without any batch-norm and with leaky Re LUs (Maas et al., 2013) instead or Re LUs. The generator takes as input a Gaussian noise vector and maps it to a 2048 3 output via 5 FC-Re LU layers.