# SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud

**Authors:** Bichen Wu, Alvin Wan, Xiangyu Yue, Kurt Keutzer  
**arXiv:** [1710.07368](https://arxiv.org/abs/1710.07368)  
**Date:** 2017-10-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1710.07368v1/

## Key points

- The paper addresses the problem of semantic segmentation of road-objects from 3D LiDAR point clouds.
- The authors propose an end-to-end pipeline called SqueezeSeg based on convolutional neural networks (CNN) that takes a transformed LiDAR point cloud as input and directly outputs a point-wise label map that is refined by a conditional random field (CRF) implemented as a recurrent layer.
- Instance-level labels are then obtained by conventional clustering algorithms.
- The authors trained their CNN model on LiDAR point clouds from the KITTI dataset and synthesized large amounts of realistic training data using a LiDAR simulator built into Grand Theft Auto V (GTA-V).
- SqueezeSeg achieves high accuracy with fast and stable runtime (8.7 ms per frame), highly desirable for autonomous driving applications.
- Additional training on synthesized data boosts validation accuracy on real-world data.
- The authors plan to open source their source code and synthesized data for others to use in future research endeavors.

## Summary

In their paper titled "SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud," Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer address the problem of semantic segmentation of road-objects from 3D LiDAR point clouds. The authors formulate this problem as a point-wise classification problem and propose an end-to-end pipeline called SqueezeSeg based on convolutional neural networks (CNN). The CNN takes a transformed LiDAR point cloud as input and directly outputs a point-wise label map that is refined by a conditional random field (CRF) implemented as a recurrent layer. Instance-level labels are then obtained by conventional clustering algorithms. The authors train their CNN model on LiDAR point clouds from the KITTI dataset, and their point-wise segmentation labels are derived from 3D bounding boxes from KITTI. To obtain extra training data, they built a LiDAR simulator into Grand Theft Auto V (GTA-V), a popular video game, to synthesize large amounts of realistic training data. Their experiments show that SqueezeSeg achieves high accuracy with astonishingly fast and stable runtime (8.7 ms per frame), highly desirable for autonomous driving applications. Furthermore, additional training on synthesized data boosts validation accuracy on real-world data. The authors plan to open source their source code and synthesized data for others to use in future research endeavors. This paper provides valuable insights into the development of real-time road object segmentation using 3D LiDAR point clouds through an end-to-end pipeline called SqueezeSeg based on CNNs with recurrent CRFs for autonomous driving applications with high accuracy and fast runtime performance.