Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving 3D Occupancy Prediction through Class-balancing Loss and Multi-scale Representation

Huizhou Chen, Jiangyi Wang, Yuxin Li, Na Zhao, Jun Cheng, Xulei Yang

TL;DR

The paper tackles 3D occupancy prediction for autonomous driving using multi-view imagery. It introduces a UNet-like Multi-scale Occupancy Head to capture scale-aware BEV features and a class-balancing loss to mitigate severe voxel-level class imbalance, both integrated with a BEVDet4D encoder. On the nuScenes occupancy dataset, the method achieves superior per-class IoU (mIoU) over baselines, particularly for rare classes, validating the value of multi-scale supervision and imbalance handling. Overall, the work enhances BEV-based occupancy prediction by enriching 3D context and balancing training signals, enabling more reliable scene understanding for planning and safety.

Abstract

3D environment recognition is essential for autonomous driving systems, as autonomous vehicles require a comprehensive understanding of surrounding scenes. Recently, the predominant approach to define this real-life problem is through 3D occupancy prediction. It attempts to predict the occupancy states and semantic labels for all voxels in 3D space, which enhances the perception capability. Birds-Eye-View(BEV)-based perception has achieved the SOTA performance for this task. Nonetheless, this architecture fails to represent various scales of BEV features. In this paper, inspired by the success of UNet in semantic segmentation tasks, we introduce a novel UNet-like Multi-scale Occupancy Head module to relieve this issue. Furthermore, we propose the class-balancing loss to compensate for rare classes in the dataset. The experimental results on nuScenes 3D occupancy challenge dataset show the superiority of our proposed approach over baseline and SOTA methods.

Improving 3D Occupancy Prediction through Class-balancing Loss and Multi-scale Representation

TL;DR

The paper tackles 3D occupancy prediction for autonomous driving using multi-view imagery. It introduces a UNet-like Multi-scale Occupancy Head to capture scale-aware BEV features and a class-balancing loss to mitigate severe voxel-level class imbalance, both integrated with a BEVDet4D encoder. On the nuScenes occupancy dataset, the method achieves superior per-class IoU (mIoU) over baselines, particularly for rare classes, validating the value of multi-scale supervision and imbalance handling. Overall, the work enhances BEV-based occupancy prediction by enriching 3D context and balancing training signals, enabling more reliable scene understanding for planning and safety.

Abstract

3D environment recognition is essential for autonomous driving systems, as autonomous vehicles require a comprehensive understanding of surrounding scenes. Recently, the predominant approach to define this real-life problem is through 3D occupancy prediction. It attempts to predict the occupancy states and semantic labels for all voxels in 3D space, which enhances the perception capability. Birds-Eye-View(BEV)-based perception has achieved the SOTA performance for this task. Nonetheless, this architecture fails to represent various scales of BEV features. In this paper, inspired by the success of UNet in semantic segmentation tasks, we introduce a novel UNet-like Multi-scale Occupancy Head module to relieve this issue. Furthermore, we propose the class-balancing loss to compensate for rare classes in the dataset. The experimental results on nuScenes 3D occupancy challenge dataset show the superiority of our proposed approach over baseline and SOTA methods.
Paper Structure (12 sections, 1 equation, 3 figures, 1 table)

This paper contains 12 sections, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Voxel-based representation
  4. BEV-based perception
  5. Methodology
  6. UNet-like Multi-scale Occupancy Head
  7. Class-balancing Loss
  8. Experiments
  9. Dataset
  10. Implementation
  11. Experimental results
  12. Conclusion

Figures (3)

  • Figure 1: Overall Architecture of BEV-based Occupancy Prediction Model.
  • Figure 2: UNet-like Multi-Scale Occupancy Head. The architecture of proposed occupancy head within one layer. The arrows, distinguished by various colors, symbolize distinct operations, while the numbers enclosed in parentheses denote the dimension of the data, batch size is omitted here.
  • Figure 3: Distribution of nuScenes Occupancy Voxels: The x-axis represents the class names, and the y-axis displays the cumulative voxel counts across all samples within the dataset.