LSCA-RCNN: Large-Kernel Spatial Residual and Cascade Attention Network for Voxel-Based 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• LiDAR-based 3D object detection faces challenges with sparse data, impacting accuracy for small/occluded objects.
• Existing methods struggle with feature degradation and limited receptive fields in complex scenes.

Purpose of the Study:

• To propose LSCA-RCNN, a novel two-stage voxel-based 3D detector.
• To enhance detection accuracy for small and occluded objects in LiDAR point clouds.
• To improve localization precision and robustness in autonomous driving.

Main Methods:

• Integrated spatial residual blocks (SRBs) and large-kernel convolutions in a 3D backbone for stable multi-scale feature learning.
• Employed a ConvNeXt-based 2D backbone with spatial attention to boost small object feature representation.
• Designed a cascaded detection head with grouped convolutions and cross-stage cross-attention for progressive refinement.

Main Results:

• LSCA-RCNN demonstrated consistent performance improvements over baseline methods on the KITTI dataset (R40 metric).
• Achieved significant 3D Average Precision (AP) gains: +2.12% (cars), +7.66% (pedestrians), +5.43% (cyclists) in moderate settings.
• +1.62% (cars), +5.05% (pedestrians), +7.05% (cyclists) gains in hard settings.

Conclusions:

• The proposed LSCA-RCNN effectively addresses challenges in LiDAR-based 3D object detection.
• The method shows significant improvements in accuracy and localization for critical objects in autonomous driving scenarios.
• LSCA-RCNN offers a robust solution for complex and demanding autonomous driving detection tasks.