SVRS: self-supervised 3D voxel reconstruction network from stereo vision.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Autonomous robots require accurate 3D environmental perception using stereo vision.
• Existing pseudo-LiDAR methods suffer from high computational costs and boundary over-smoothing in voxel grid reconstruction.

Purpose of the Study:

• To develop an efficient and accurate 3D voxel reconstruction method from stereo vision for autonomous robots.
• To overcome the limitations of existing pseudo-LiDAR approaches.

Main Methods:

• Proposed a Self-supervised 3D Voxel Reconstruction network from Stereo vision (SVRS).
• Introduced a Pixel-Voxel Projecting Module (PVPM) to establish stereo-voxel correspondences and convert dense pixel features to sparse voxel representations.
• Utilized an Octree-based Encoder-Decoder Architecture (OEDA) for hierarchical, multi-scale voxel grid reconstruction.
• Employed a self-supervised training framework with off-the-shelf stereo matching methods.

Main Results:

• SVRS achieves competitive reconstruction accuracy on the DrivingStereo dataset.
• SVRS demonstrates significant improvements in inference speed, up to 14x faster than advanced pseudo-LiDAR methods and 3x faster than real-time approaches.

Conclusions:

• SVRS offers an efficient and accurate solution for 3D voxel reconstruction from stereo vision.
• The proposed method enhances environmental perception capabilities for autonomous robots.