EGOF-Net: epipolar guided optical flow network for unrectified stereo matching.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Dynamic stereo vision systems face challenges in accurate stereo matching.
• Existing methods often require rectification, complicating the pipeline.

Purpose of the Study:

• To develop a novel network for robust unrectified stereo matching in dynamic environments.
• To improve the accuracy and efficiency of 3D point cloud reconstruction.

Main Methods:

• Proposed an epipolar guided optical flow network (EGOF-Net).
• Introduced a deep cross-checking-based fundamental matrix estimation method (DCCM) for robust epipolar geometry.
• Implemented a 4D epipolar modulator (4D-EM) to eliminate false matches.

Main Results:

• EGOF-Net significantly outperforms state-of-the-art methods on synthetic and real-scene datasets.
• Successfully reconstructed 3D point clouds in an existing dynamic stereo system.
• Demonstrated the ability to bypass rectification operations.

Conclusions:

• The proposed EGOF-Net offers a robust solution for unrectified stereo matching in dynamic systems.
• The method simplifies the stereo vision pipeline and enables new heuristic-neural network combinations.
• The approach shows promise for real-world dynamic stereo vision applications.