# A General Pipeline for 3D Detection of Vehicles

**Authors:** Xinxin Du, Marcelo H. Ang Jr., Sertac Karaman, Daniela Rus  
**arXiv:** [1803.00387](https://arxiv.org/abs/1803.00387)  
**Date:** 2018-02-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1803.00387v1/

## Key points

- The paper addresses the need for accurate 3D perception in autonomous driving systems
- Proposes a flexible pipeline that can incorporate any 2D detection network and fuse it with a 3D point cloud
- Develops an effective model fitting algorithm based on generalized car models and score maps to identify the 3D box
- Proposes a two-stage convolutional neural network (CNN) to refine the detected 3D box further
- Conducts experiments using two different 2D detection networks on the KITTI dataset, yielding similar 3D detection outcomes
- Ranks second overall when compared against other state-of-the-art algorithms for 3D detection
- Presents a general pipeline that integrates existing 2D detection networks with a 3D point cloud for robust and accurate vehicle detection in autonomous driving systems

## Summary

The paper titled "A General Pipeline for 3D Detection of Vehicles" addresses the need for accurate 3D perception in autonomous driving systems. While current methods primarily focus on 2D vehicle detection, this paper proposes a flexible pipeline that can incorporate any 2D detection network and fuse it with a 3D point cloud to generate comprehensive 3D information with minimal modifications to the existing networks. To achieve this, the authors develop an effective model fitting algorithm based on generalized car models and score maps to identify the 3D box. Additionally, they propose a two-stage convolutional neural network (CNN) to refine the detected 3D box further. To evaluate the performance of their pipeline, the authors conduct experiments using two different 2D detection networks on the KITTI dataset. The results demonstrate that both networks yield similar 3D detection outcomes, highlighting the flexibility of the proposed pipeline. Furthermore, when compared against other state-of-the-art algorithms for 3D detection, this pipeline achieves impressive results and ranks second overall. This outcome underscores its competence in accurately detecting vehicles in a three-dimensional space. In conclusion, this paper presents a general pipeline that effectively integrates existing 2D detection networks with a 3D point cloud to enable robust and accurate vehicle detection in autonomous driving systems. Its flexibility and high-performance capabilities make it a promising solution for advancing real-world applications of autonomous vehicles.