A Survey on Deep-Learning-Based LiDAR 3D Object Detection for Autonomous Driving

Round 1

Reviewer 1 Report

I appreciate the efforts the authors have made. This paper is well-organized and written. I only two comments before acceptance

1) Please point out the conflicts and deficiencies of exisiting review works. This makes reader better understand the merit of this paper.

2) Please cite the following paper: Large-Scale Bandwidth and Power Optimization for Multi-Modal Edge Intelligence Autonomous Driving 

Author Response

Point 1: Please point out the conflicts and deficiencies of existing review works. This makes the reader better understand the merit of this paper.

Response 1: We want to thank you for your insightful comments on improving the manuscript. Most of the existing 3D object detection methods reviewed general 2D and 3D detection methods and/or LiDAR and camera-based detections. This survey focuses detailed analysis of LiDAR-only methods. Few works reviewed LiDAR-only methods. However, we have included recently published works in addition to unique contributions, such as 3D coordinate systems in 3D object detection and stages of autonomous driving. We updated the related section by adding weaknesses of existing review works and unique contributions of this survey work. The contribution section also shows the contribution of this survey paper.

Point 2: Please cite the following paper: Large-Scale Bandwidth and Power Optimization for Multi-Modal Edge Intelligence Autonomous Driving.

Response 2: Thank you for suggesting the paper. We included the suggested work to strengthen our manuscript.

Reviewer 2 Report

Target detection based on point clouds  is a hot spot in current research, and the paper review is relatively comprehensive and of great reference value. The main recommendations are as follows：

1） Fig. 4 and Fig. 5 can refer directly to Figures 1 and 2，

2）The target detection of point clouds is not limited to autonomous driving applications, and the paper emphasizes that the cause of  autonomous driving is unknown; On the contrary, autonomous  driving requires algorithms to have high real-time and accuracy, which is not reflected in the paper.

Author Response

Point 1: Fig. 4 and Fig. 5 can refer directly to Figures 1 and 2，

Response 1: We want to thank you for your insightful comments on improving the manuscript. Yes, Figure 4 is similar to Figure1, but Figure 2 and Figure 5 are different. We modified the manuscript based on the comment.  

Point 2: The target detection of point clouds is not limited to autonomous driving applications, and the paper emphasizes that the cause of autonomous driving is unknown; On the contrary, autonomous driving requires algorithms to have high real-time and accuracy, which is not reflected in the paper.

Response 2: The application of LiDAR ranges from geospatial mapping, forestry, and smart city to scene understanding in autonomous driving. This survey focuses only on LiDAR-based 3D object detection methods, LiDAR processing techniques, and others for autonomous driving. The LiDAR processing techniques presented in this survey can be used in other applications. As stated in the introduction section, autonomous driving needs a robust model to understand the driving environment under different weather, fast decision-making system during high-speed driving, and accurate information about the driving environment. Most current 3D perception system models, especially 3D object detection, need to be more lightweight to apply for real-time deployment as of 2D equivalents. More work and innovations are required to achieve level four and five driving, which have yet to be commercially available. However, we summarized state-of-the-art 3D object detection methods for autonomous driving.