# Complex-YOLO: Real-time 3D Object Detection on Point Clouds

**Authors:** Martin Simon, Stefan Milz, Karl Amende, Horst-Michael Gross  
**arXiv:** [1803.06199](https://arxiv.org/abs/1803.06199)  
**Date:** 2018-03-16  
**Summary page:** https://www.summarizepaper.com/en/paper/1803.06199v1/

## Key points

- Lidar-based 3D object detection is crucial for autonomous driving and other areas
- Real-time inference of sparse 3D data is a challenge in various fields
- Complex-YOLO is a state-of-the-art real-time 3D object detection network for point clouds
- It expands upon YOLOv2 by incorporating complex regression strategy and Euler Region Proposal Network (E-RPN)
- Complex-YOLO outperforms current leading methods in terms of efficiency and accuracy
- Achieves state-of-the-art results for various classes including cars, pedestrians, and cyclists
- More than five times faster than its closest competitor
- High accuracy in estimating all eight KITTI classes simultaneously
- Significant implications for autonomous driving, augmented reality, personal robotics, and industrial automation.

## Summary

The paper titled "Complex-YOLO: Real-time 3D Object Detection on Point Clouds" by Martin Simon, Stefan Milz, Karl Amende, and Horst-Michael Gross focuses on the importance of lidar-based 3D object detection for autonomous driving. The authors highlight that this technology is crucial for environmental understanding and forms the foundation for prediction and motion planning in autonomous vehicles. However, the real-time inference of highly sparse 3D data poses a challenge not only in automated vehicles but also in other areas such as augmented reality, personal robotics, and industrial automation. To address this challenge, the authors introduce Complex-YOLO, a state-of-the-art real-time 3D object detection network that operates solely on point clouds. They expand upon YOLOv2 - a fast 2D standard object detector for RGB images - by incorporating a specific complex regression strategy to estimate multi-class 3D boxes in Cartesian space. This involves proposing an Euler Region Proposal Network (E-RPN) that estimates the pose of objects by adding imaginary and real fractions to the regression network. By operating in a closed complex space and avoiding singularities caused by single angle estimations, the E-RPN ensures robustness during training. The authors conduct experiments using the KITTI benchmark suite to evaluate their proposed method's performance. The results demonstrate that Complex-YOLO outperforms current leading methods for 3D object detection in terms of efficiency. It achieves state-of-the art results for various classes including cars, pedestrians and cyclists while being more than five times faster than its closest competitor. Furthermore, Complex YOLO exhibits high accuracy in estimating all eight KITTI classes simultaneously including Vans Trucks and sitting pedestrians. In conclusion, this paper presents Complex YOLO as an effective solution for real time 3D object detection on point clouds. Its innovative approach expands upon existing techniques and achieves superior performance in terms of efficiency and accuracy. The findings have significant implications not only for autonomous driving but also for other domains such as augmented reality personal robotics and industrial automation.