MsSVT++: Mixed-Scale Sparse Voxel Transformer With Center Voting for 3D Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Accurate 3D object detection in large-scale outdoor environments is crucial for applications like autonomous driving.
• Existing methods using window-based transformers struggle to balance long-range dependencies with fine-grained details.

Purpose of the Study:

• To develop a novel transformer architecture for 3D object detection that captures both long-range and fine-grained features.
• To improve the accuracy and efficiency of 3D object detection in complex outdoor scenes.

Main Methods:

• Proposed MsSVT++ (Mixed-scale Sparse Voxel Transformer) with a divide-and-conquer attention mechanism.
• Introduced Chessboard Sampling and sparse voxel operations using a hash map to reduce computational complexity.
• Developed a Center Voting module to enhance bounding box localization by integrating contextual information.

Main Results:

• MsSVT++ effectively integrates mixed-scale features, addressing limitations of previous approaches.
• The Chessboard Sampling and sparse operations significantly improve computational efficiency.
• The Center Voting module demonstrably enhances the precision of object localization.

Conclusions:

• MsSVT++ provides a robust and efficient solution for 3D object detection in large-scale outdoor scenes.
• The proposed methods significantly advance the state-of-the-art in accurate and detailed 3D object recognition.