IGEV++: Iterative Multi-Range Geometry Encoding Volumes for Stereo Matching.

Area of Science:

• Computer Vision
• Robotics
• Deep Learning

Background:

• Stereo matching is crucial for computer vision and robotics.
• Handling ambiguities in ill-posed regions and large disparities remains a significant challenge.

Purpose of the Study:

• To propose IGEV++, a novel deep network architecture for improved stereo matching.
• To address limitations in current stereo matching methods, particularly in challenging scenarios.

Main Methods:

• Introduction of Multi-range Geometry Encoding Volumes (MGEV) to encode multi-granularity geometry information.
• Development of an adaptive patch matching module for efficient cost computation in large disparity ranges.
• Utilization of a selective geometry feature fusion module and ConvGRUs for iterative disparity map refinement.

Main Results:

• IGEV++ achieves state-of-the-art performance on Scene Flow, Middlebury, ETH3D, and KITTI benchmarks.
• Demonstrates significant error reduction on large disparity benchmarks compared to existing methods.
• A real-time version of IGEV++ achieves top performance among published real-time stereo matching methods.

Conclusions:

• IGEV++ effectively handles large disparities and ill-posed regions, leading to superior stereo matching accuracy.
• The proposed MGEV approach offers efficient processing and rapid convergence.
• IGEV++ represents a significant advancement in stereo matching technology for computer vision and robotics applications.