AGO-Net: Association-Guided 3D Point Cloud Object Detection Network.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Current 3D object detection from LiDAR struggles with occluded and distant objects due to varying point cloud appearance and density.
• Robust feature representations are critical for overcoming these challenges in 3D object detection.

Purpose of the Study:

• To develop a novel 3D object detection framework that enhances feature robustness for occluded and distant objects.
• To bridge the gap between real-world perceptual data and augmented conceptual data for improved 3D object detection.

Main Methods:

• Proposed a 3D detection framework utilizing domain adaptation to associate intact object features.
• Constructed conceptual scenes from augmented data without external datasets.
• Introduced an attention-based re-weighting module for adaptive feature adaptation.

Main Results:

• Achieved state-of-the-art performance on the KITTI 3D detection benchmark in both accuracy and speed.
• Demonstrated the method's versatility and effectiveness on nuScenes and Waymo datasets.
• The attention module enhances feature adaptation without increasing inference costs.

Conclusions:

• The proposed domain adaptation framework significantly improves 3D object detection for challenging scenarios.
• The plug-and-play attention module offers a versatile solution for various 3D detection systems.
• This approach advances the capabilities of LiDAR-based 3D object detection.