DALI: Domain Adaptive LiDAR Object Detection via Distribution-Level and Instance-Level Pseudolabel Denoising

Object detection using LiDAR point clouds relies on a large amount of human-annotated samples when training the underlying detectors' deep neural networks. However, generating 3-D bounding box annotation for a large-scale dataset could be costly and time-consuming. Alternatively, unsupervised domain adaptation (UDA) enables a given object detector to operate on novel new data, with an unlabeled training dataset, by transferring the knowledge learned from training labeled
source domain
data to the new unlabeled
target domain
. Pseudolabel strategies, which involve training the 3-D object detector using target-domain predicted bounding boxes from a pretrained model, are commonly used in UDA. However, these pseudolabels often introduce noise, impacting performance. In this article, we introduce the domain adaptive LiDAR (DALI) object detection framework to address noise at both distribution and instance levels. First, a posttraining size normalization (PTSN) strategy is developed to mitigate bias in pseudolabel size distribution by identifying an unbiased scale after network training. To address instance-level noise between pseudolabels and corresponding point clouds, two pseudopoint clouds generation (PPCG) strategies, ray-constrained and constraint-free, are developed to generate pseudopoint clouds for each instance, ensuring the consistency between pseudolabels and pseudopoints during training. We demonstrate the effectiveness of our method on the publicly available and popular datasets KITTI, Waymo, and nuScenes. We show that the proposed DALI framework achieves state-of-the-art results and outperforms leading approaches on most of the domain adaptation tasks.