HBEVOcc: Height-Aware Bird's-Eye-View Representation for 3D Occupancy Prediction from Multi-Camera Images.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• 3D occupancy prediction is vital for autonomous driving and robotics, enabling detailed scene perception and object recognition.
• Voxel-based methods dominate 3D occupancy prediction but suffer from high memory and computational costs.
• Efficient 3D scene representation is crucial for advancing vision-centric autonomous systems.

Purpose of the Study:

• To introduce HBEVOcc, an efficient Bird's-Eye-View (BEV) based method for 3D occupancy prediction.
• To address the memory and computational limitations of existing voxel-based approaches.
• To enhance 3D scene representation by effectively utilizing latent height information.

Main Methods:

• HBEVOcc employs a novel height-aware deformable attention module within a BEV framework.
• It extracts multi-camera image features and transforms them into 3D BEV occupancy features.
• A height-aware voxel loss with adaptive vertical weighting is introduced for improved supervision.

Main Results:

• HBEVOcc achieves state-of-the-art performance on Occ3D-nuScenes and OpenOcc datasets.
• The method demonstrates significant reductions in training memory requirements, even on consumer-grade hardware (2080Ti).
• Performance gains are observed in both mIoU and RayIoU metrics, indicating enhanced 3D occupancy prediction accuracy.

Conclusions:

• HBEVOcc offers a computationally efficient and high-performing solution for 3D occupancy prediction.
• The height-aware deformable attention module effectively compensates for the lack of explicit height information in BEV representations.
• This approach facilitates the development of more resource-efficient and capable autonomous driving and robotics systems.