PseudoAugment: Learning to Use Unlabeled Data for Data Augmentation in Point Clouds

Data augmentation is an important technique to improve data efficiency and save labeling cost for 3D detection in point clouds. Yet, existing policies have so far been designed only utilize labeled data, which limits the diversity. In this paper, we recognize that pseudo are complementary, thus propose leverage unlabeled enrich training data. particular, design three novel pseudo-label based (PseudoAugments) fuse both pseudo-labeled scenes, including frames (PseudoFrame), objecta (PseudoBBox), background (PseudoBackground). PseudoAugments outperforms by mitigating errors generating diverse fused scenes. We demonstrate generalize across point-based voxel-based architectures, different model capacity KITTI Waymo Open Dataset. To alleviate of hyperparameter tuning iterative labeling, develop a population-based framework detection, named AutoPseudoAugment. Unlike previous works perform pseudo-labeling offline, our performs one shot reduce computational cost. Experimental results on large-scale Dataset show method state-of-the-art auto (PPBA) self-training (pseudo labeling). AutoPseudoAugment about 3X 2X efficient vehicle pedestrian tasks compared prior arts. Notably, nearly matches full dataset results, with just 10% run segments task.