# BEV-SAN: Accurate BEV 3D Object Detection via Slice Attention Networks

**Authors:** Xiaowei Chi, Jiaming Liu, Ming Lu, Rongyu Zhang, Zhaoqing Wang, Yandong Guo, Shanghang Zhang  
**arXiv:** [2212.01231](https://arxiv.org/abs/2212.01231)  
**Date:** 2022-12-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.01231v1/

## Key points

- Autonomous driving systems play a pivotal role in ensuring vehicle safety and efficiency on the road.
- Recent advancements in autonomous driving systems focus on three key components: camera feature extraction, BEV feature construction, and task heads.
- BEV feature construction presents unique challenges compared to traditional 2D tasks, particularly in capturing informative features at different heights within the BEV space.
- A novel approach called [novel approach name] introduces a new sampling technique along the height dimension to create global and local BEV slices for accurate object detection in 3D space.
- The inclusion of local BEV slices highlights specific height-related details crucial for accurate object detection by leveraging statistical distributions obtained from LiDAR data.
- Extensive experiments have validated the efficacy of this approach in accurately detecting objects in complex 3D environments within autonomous driving systems.

## Summary

In the realm of autonomous driving systems,  plays a pivotal role in ensuring the safety and efficiency of vehicles on the road. Recent advancements in this field have seen the emergence of various works that adhere to a common framework comprising three key components: camera feature extraction, BEV feature construction, and task heads. While these components are essential for accurate object detection in a 3D space, it is noted that BEV feature construction presents unique challenges compared to traditional 2D tasks. One particular challenge lies in effectively capturing and emphasizing informative features at different heights within the BEV space. For instance, objects like barriers may be situated at lower heights while larger vehicles such as trucks may occupy higher positions. To address this issue, a novel approach known as  has been proposed. Unlike conventional methods that flatten the BEV space along the height dimension,  introduces a new sampling technique along the height dimension to create both global and local BEV slices. By aggregating features from these distinct BEV slices derived from camera inputs and leveraging an attention mechanism for merging them,  aims to enhance the representation of informative heights within the 3D space. Furthermore, a transformer-based fusion process is employed to combine the merged local and global features into a final feature map for task heads. Notably, the inclusion of local BEV slices serves to highlight specific height-related details crucial for accurate object detection. To determine which heights should be emphasized in the local slices, a  is proposed based on statistical distributions obtained from LiDAR data. This strategy enables more informed selection of heights compared to uniform sampling methods, ultimately enhancing the effectiveness of object detection algorithms. Extensive experiments have been conducted to validate the efficacy of  in accurately detecting objects in complex 3D environments. Additionally, it is worth noting that code associated with this innovative approach will be made publicly available for further research and development efforts. In summary,  represents a significant advancement in addressing height-related challenges inherent in 3D object detection tasks within autonomous driving systems. By leveraging advanced sampling techniques and attention mechanisms, this approach offers promise for improving accuracy and reliability in identifying objects across varying heights within the BEV space.