.png){kind=icon}
.png){kind=icon}
Return
GLA: Globally-aware graph and dilated local attention for enhanced 3D object detection
.png){kind=icon}
PRE
AI
L
L
S
H
DOI：10.1016/j.patrec.2026.03.013.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recently, deep learning-based 3D object detection methods have achieved remarkable progress in complex indoor and outdoor environments. However, existing approaches still face challenges in effectively modeling global contextual information and extracting fine-grained features, particularly when detecting distant or boundary-blurred objects. To address these limitations, we propose an enhanced 3D object detection framework named Global-Local Attention Network (GLA), built upon the PV-RCNN architecture. GLA introduces two key modules Globally-Aware Graph Convolution (GAGC) and Multi-Scale Dilated Local Attention (MDLA)strengthen global semantic reasoning and local feature representation. Specifically, GAGC leverages a dual-scale KNN graph construction mechanism to capture long-range dependencies in point clouds, while MDLA integrates depthwise separable convolutions with local attention to improve the perception of small objects and local geometry. Experiments on KITTI benchmark show that gla has achieved competitive performance in multiple categories and difficulty levels, and has obvious advantages compared with the new mainstream methods, especially in detecting longdistance and small-scale objects. These results verify that GLA achieves a good balance between accuracy and computational efficiency.
Keywords:
3D object detection
Autonomous driving
Point cloud
Deep learning
Journal
IF:
3.3
Papers:
7.8K
Citations:
1.6W
Organization
Cited Papers