Vision BEV-MAE: Masked Bird's Eye View Autoencoders for Camera-based 3D Object Detection

3D object detection is one of the fundamental per-ception tasks in autonomous driving. In recent years, perception in the Bird's Eye View (BEV) has attracted tremendous attention due to its outstanding ability in preserving geometric cues. How-ever, like other supervised learning tasks, BEV-based perception suffers from inadequate labeled training data. In this work, we propose the Vision BEV-MAE, a self-supervised learning approach for BEV perception from camera images. Specifically, we design a feature-level masking strategy to mask the mid-level representations, i.e., the encoded BEV features instead of directly masking input multi-view images, and reconstruct BEV features with a light-weight decoder. Our Vision BEV-MAE is independent of any geometric data such as point clouds, or any additional supervision, and can be easily implemented within any state-of-the-art BEV perception frameworks to improve the performance. We conduct experiments on benchmark dataset to validate our approach.