.png){kind=icon}
.png){kind=icon}
Return
Shape Enhanced Network With Transformer-Based Completion for LiDAR-Based 3D Detection From Point Cloud
.png){kind=icon}
PRE
AI
DOI：10.1109/TASE.2026.3652601.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In LiDAR-based 3D object detection, severe occlusion often leads to incomplete object shape in point cloud scene. It is a critical challenge as it will significantly degrade detection performance, ultimately jeopardizing the safety of autonomous driving system. The local shape completion technology of incomplete point cloud is therefore paramount for achieving robust perception in real 3D scenarios. However, most previous works overlook this important challenge. To tackle this issue, we present a shape enhanced 3D detection network named SET-Det under obstructed environments in this paper. Specifically, SET-Det contains a 3D proposal prediction stage with Region Proposal Network (RPN) and a completion refinement stage with a few 3D proposals and incomplete partial points as inputs. The designed shape completion model utilizes a Transformer-based feature encoder to model the correlation among input partial points, and restores the complete geometric shapes via a shape generation decoder stacked several Multi-layer Perceptrons (MLPs) and fully connected layers. Relevant experiments on KITTI and Waymo Open datasets demonstrate our SET-Det achieves state-of-the-art performance. Specifically, on the KITTI benchmark, SET-Det improves the mAP by 0.88% and 1.83% for car class on validation and testing sets compared to baseline Voxel-RCNN. More importantly, it shows significant performance gains of 2.16% and 13.57% for pedestrian and cyclist under hard occlusion scenarios on KITTI validation set in comparison of multi-modal method EPNet++, validating its effectiveness in handling the challenging occluded point cloud problem. Note to Practitioners—The motivation of this work is to address the incomplete point cloud shape scanned by LiDAR in 3D detection task for autonomous driving. In actual outdoor environments, due to occlusion, long distance, and limitations of certain reflection angles, LiDAR cannot obtain complete point cloud which describes the object shape information, thereby affecting the final 3D detection performance. However, most previous LiDAR-based 3D detection methods overlook this challenge. Therefore, this paper proposes a shape enhanced 3D detection network named SET-Det under obstructed environments. Specifically, we design a shape completion model containing a Transformer-based encoder and a shape generation decoder to restore the complete geometric shapes of detected objects. Relevant experiments on two open outdoor datasets verify the effectiveness of our proposed method. The proposed method can be deployed for 3D perception task in autonomous driving, as well as other LiDAR-based detection and recognition tasks.
Keywords:
3D object detection
point cloud
autonomous driving
shape completion
obstructed environments
Journal
IF:
6.4
Papers:
4.9K
Citations:
1.6W