Vicente Muñoz's Thesis Presentation

Title: Use of Transformers for Automatic Symmetry Plane Detection in 3D Point Clouds

Abstract: This proposal aims to improve the automatic detection of symmetry planes in 3D point clouds using Transformers—specifically, a scalable Point Cloud Transformer—and to compare its performance with reference ConvNet architectures (PointNet++, DGCNN, SymmetryNet) using the Symmetria dataset, a synthetic dataset comprising four difficulty levels with increasing noise, sub-sampling, and rotations. The plan includes replicating existing baselines, training the new model on GPU servers at the University of Chile, evaluating its accuracy (mAP, PHC) in easy→hard, zero-shot, and few-shot scenarios, and analyzing how increasing parameters and data aligns with the scaling law proposed by Kaplan et al., with the expectation that the Transformer will outperform other methods in both robustness and accuracy.

Details

• Room: Ada Lovelace, West Building
• Date: Friday, August 7th at 11:00 AM