Five aspects of researches about automatic drive perception by Waymo_.rsn: range sparse net for efficient, accurate lid-优快云博客

本文链接：https://blog.youkuaiyun.com/weixin_46187561/article/details/121879797

本文汇总了多项关于3D对象检测的研究成果，包括Pointpillar的快速点云编码器、PV-RCNN++的点-体素特征抽象、RSN的范围稀疏网络以及3D-Man的多帧注意力网络。这些工作通过创新的方法提升了激光雷达输入质量，利用多帧信息进行互补融合，并实现了高效准确的检测。此外，还介绍了SurfelGAN生成逼真的传感器数据以助力自动驾驶研究。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Main stream network model:

Pointpillar: Fast Encoders for Object Detection from Point Clouds by Lang et al.

paper:https://arxiv.org/pdf/1812.05784.pdf

PV-RCNN++:Point-Voxel Feature Set Abstraction for 3D Object Detection by Shi et al.

paper:https://arxiv.org/pdf/2102.00463.pdf

Five aspects of researches about automatic drive perception by Waymo.

SPG：Unsupervised Domain Adaptation for 3D Object Detection via Semantic Point Generation

targeted & light-weight model for improving LIDAR input quality against occlusions and poor weathers
Effectiveness verified by applying to popular and SOTA 3D detectors.
complete the 3D shape before Detection

paper:https://openaccess.thecvf.com/content/ICCV2021/papers/Xu_SPG_Unsupervised_Domain_Adaptation_for_3D_Object_Detection_via_Semantic_ICCV_2021_paper.pdf

3D-man：3D Multi-frame Attention Network for Object Detection

Complementary information from multiple frames
learned attention to fuse multi-frame
learn to fuse complementary information from multi-frame data via attention layers

paper:https://openaccess.thecvf.com/content/CVPR2021/papers/Yang_3D-MAN_3D_Multi-Frame_Attention_Network_for_Object_Detection_CVPR_2021_paper.pdf

RSN: Range Sparse Net for Efficient,Accurate LIDAR 3D Object Detection

targeted feature selection via segmentation
Efficient backbone based on sparse conv

paper:https://openaccess.thecvf.com/content/CVPR2021/papers/Sun_RSN_Range_Sparse_Net_for_Efficient_Accurate_LiDAR_3D_Object_CVPR_2021_paper.pdf

Labeling Automation: Offboard 3D Object Detection from Point Cloud Sequences

leverage structured information in the 3D space and temporal sequences.
Quality on-par with human labelers

paper:https://openaccess.thecvf.com/content/CVPR2021/papers/Qi_Offboard_3D_Object_Detection_From_Point_Cloud_Sequences_CVPR_2021_paper.pdf

SurfelGAN: Synthesizing Realistic Sensor Data for Autonomous Driving

Generating realistic camera images for novel SDC and dynamic object poses
Scalable solution based on pre-collected LIDAR and camera data

paper:https://openaccess.thecvf.com/content_CVPR_2020/papers/Yang_SurfelGAN_Synthesizing_Realistic_Sensor_Data_for_Autonomous_Driving_CVPR_2020_paper.pdf

Five aspects of researches about automatic drive perception by Waymo

Main stream network model:

Pointpillar: Fast Encoders for Object Detection from Point Clouds by Lang et al.

paper:https://arxiv.org/pdf/1812.05784.pdf

PV-RCNN++:Point-Voxel Feature Set Abstraction for 3D Object Detection by Shi et al.

paper:https://arxiv.org/pdf/2102.00463.pdf

Five aspects of researches about automatic drive perception by Waymo.

SPG：Unsupervised Domain Adaptation for 3D Object Detection via Semantic Point Generation

targeted & light-weight model for improving LIDAR input quality against occlusions and poor weathers

Effectiveness verified by applying to popular and SOTA 3D detectors.

complete the 3D shape before Detection

paper:https://openaccess.thecvf.com/content/ICCV2021/papers/Xu_SPG_Unsupervised_Domain_Adaptation_for_3D_Object_Detection_via_Semantic_ICCV_2021_paper.pdf

3D-man：3D Multi-frame Attention Network for Object Detection

Complementary information from multiple frames

learned attention to fuse multi-frame

learn to fuse complementary information from multi-frame data via attention layers

paper:https://openaccess.thecvf.com/content/CVPR2021/papers/Yang_3D-MAN_3D_Multi-Frame_Attention_Network_for_Object_Detection_CVPR_2021_paper.pdf

RSN: Range Sparse Net for Efficient,Accurate LIDAR 3D Object Detection

targeted feature selection via segmentation

Efficient backbone based on sparse conv

paper:https://openaccess.thecvf.com/content/CVPR2021/papers/Sun_RSN_Range_Sparse_Net_for_Efficient_Accurate_LiDAR_3D_Object_CVPR_2021_paper.pdf

Labeling Automation: Offboard 3D Object Detection from Point Cloud Sequences

leverage structured information in the 3D space and temporal sequences.

Quality on-par with human labelers

paper:https://openaccess.thecvf.com/content/CVPR2021/papers/Qi_Offboard_3D_Object_Detection_From_Point_Cloud_Sequences_CVPR_2021_paper.pdf

SurfelGAN: Synthesizing Realistic Sensor Data for Autonomous Driving

Generating realistic camera images for novel SDC and dynamic object poses

Scalable solution based on pre-collected LIDAR and camera data

paper:https://openaccess.thecvf.com/content_CVPR_2020/papers/Yang_SurfelGAN_Synthesizing_Realistic_Sensor_Data_for_Autonomous_Driving_CVPR_2020_paper.pdf