Graph Convolutional Neural Networks for 3D Inference

Event Sponsor: 
Argonne Leadership Computing Facility Seminar
Start Date: 
Oct 5 2018 - 10:30am
Building/Room: 
Building 240/Room 4301
Location: 
Argonne National Laboratory
Speaker(s): 
Miguel Dominguez
Speaker(s) Title: 
Rochester Institute of Technology
Host: 
Taylor Childers

Deep Neural Networks have achieved impressive performance on computer vision tasks such as classification, localization, segmentation, captioning, and generation. Applying the same techniques to 3D point clouds requires a different approach, because they are not aligned along a grid that a fixed-size kernel can slide across. We propose treating these point clouds as graphs with connections between nearest neighbors. With this structure we can define a graph convolution with a fixed-size kernel that can handle variable-size neighborhoods as well as an algebraic graph pooling o peration based on graph clustering. With these operations we build convolutional neural networks for 3D object classification on the ModelNet dataset, achieving results comparable to the state of the art. We also discuss how sparse implementations of these operations reduce their memory and computational complexity.

Miscellaneous Information: 

This seminar will be streamed, see details at https://anlpress.cels.anl.gov/mcs-streaming-seminars.

Please click here [schedule.ics] to add this event to your calendar.