Reproduce “PseudoAugment: Learning to Use Unlabeled Data for Data Augmentation in Point Clouds” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhaoqi Leng, Shuyang Cheng, Benjamin Caine, Weiyue Wang, Xiao Zhang, Jonathon Shlens, Mingxing Tan, Dragomir Anguelov

Researchers at Waymo and Google developed PseudoAugment and AutoPseudoAugment, a framework that employs novel augmentation policies to effectively leverage unlabeled 3D point cloud data for training object detectors. The method achieved nearly 10x data efficiency for vehicle detection, matching models trained on 100% labeled data using only 10% of the available labels.

Key insights

  • Fusing compositional elements (foreground objects, background) from both labeled and unlabeled 3D scenes combinatorially increases data diversity, leading to more robust 3D object detectors.
  • Aggressively mitigating pseudo-labeling errors by not only filtering labels but also removing associated point clouds is crucial for effectively leveraging unlabeled data in semi-supervised learning.
  • Integrating automated augmentation search and self-training into a single, efficient framework via population-based distillation creates a virtuous cycle that simultaneously improves pseudo-labels and model performance without prohibitive computational costs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.