SASAN: Shape-Adaptive Set Abstraction Network for Point-Voxel 3D Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving

Background:

• Point-voxel 3D object detectors show promise but struggle with varying object geometries due to fixed receptive fields and feature extraction methods.
• Existing methods lack adaptivity in modeling diverse object shapes and sizes, limiting performance in complex traffic scenarios.

Purpose of the Study:

• To propose a Shape-Adaptive Set Abstraction Network (SASAN) for enhanced point-voxel 3D object detection.
• To improve the modeling of geometrical deformations and object sizes in 3D object detection.

Main Methods:

• Developed a proposal and offset generation module for learning 3D proposals and shape-adaptive offsets.
• Introduced a shape-adaptive set abstraction module for extracting multiscale keypoint features.
• Utilized a region of interest (RoI)-grid proposal refinement module for feature aggregation and prediction.

Main Results:

• The proposed SASAN demonstrates superior performance on the KITTI 3D detection benchmark.
• Achieved significant improvements compared to existing state-of-the-art 3D object detection methods.
• The offset supervision task effectively guided the network to adapt to various object shapes.

Conclusions:

• SASAN effectively addresses the limitations of fixed receptive fields in 3D object detection.
• The shape-adaptive approach significantly enhances the detection of objects with diverse geometrical properties.
• SASAN represents a promising advancement for 3D object detection in autonomous driving applications.