EPiC: Ensemble of Partial Point Clouds for Robust Classification
ICCV 2023
ArXiv
Demo
Abstract
Point-cloud networks struggle with robustness under partiality and noise. We introduce EPiC, an ensemble method where each classifier operates on a different partial sampling—patches, curves, or random subsets—yielding diverse yet complementary predictions. This diversity acts as a consensus filter against corruptions, significantly improving robustness on ModelNet-C. EPiC achieves state-of-the-art mCE scores while remaining lightweight, backbone-agnostic, and easy to deploy.
Key Insights
Partial-Input Ensembles
Each classifier sees only a partial view of the point cloud—patches, curves, or random subsets—yielding diverse perspectives on the same object.
Diversity as Robustness
Diverse errors across partial samplings act like a noise filter: when averaged, they cancel corruptions and preserve true structure.
Consensus Mechanism
Simple averaging across ensemble members yields strong consensus, suppressing corruption effects without complex training tricks.
Practical Benefits
EPiC achieves state-of-the-art robustness on ModelNet-C while remaining lightweight, backbone-agnostic, and easy to deploy in practice.
Presentation
5 minutes presentation
BibTeX
@InProceedings{Levi_2023_ICCV,
author = {Levi, Meir Yossef and Gilboa, Guy},
title = {EPiC: Ensemble of Partial Point Clouds for Robust Classification},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2023},
pages = {14475-14484}
}