Masked Autoencoder for Self-Supervised Pre-training on Lidar Point Clouds
This work addresses the challenge of efficient 3D object detection for autonomous vehicles by enabling better pre-training with less labeled data, though it is incremental as it adapts an existing paradigm to a new domain.
The paper tackles the problem of self-supervised pre-training for lidar point clouds in automotive settings, which are sparse and have varying point densities, by proposing Voxel-MAE, a masked autoencoding scheme. The result is a 1.75 mAP and 1.05 NDS improvement on the nuScenes dataset and a 40% reduction in annotated data needed to outperform random initialization.
Masked autoencoding has become a successful pretraining paradigm for Transformer models for text, images, and, recently, point clouds. Raw automotive datasets are suitable candidates for self-supervised pre-training as they generally are cheap to collect compared to annotations for tasks like 3D object detection (OD). However, the development of masked autoencoders for point clouds has focused solely on synthetic and indoor data. Consequently, existing methods have tailored their representations and models toward small and dense point clouds with homogeneous point densities. In this work, we study masked autoencoding for point clouds in an automotive setting, which are sparse and for which the point density can vary drastically among objects in the same scene. To this end, we propose Voxel-MAE, a simple masked autoencoding pre-training scheme designed for voxel representations. We pre-train the backbone of a Transformer-based 3D object detector to reconstruct masked voxels and to distinguish between empty and non-empty voxels. Our method improves the 3D OD performance by 1.75 mAP points and 1.05 NDS on the challenging nuScenes dataset. Further, we show that by pre-training with Voxel-MAE, we require only 40% of the annotated data to outperform a randomly initialized equivalent. Code available at https://github.com/georghess/voxel-mae