Hybrid Deep-Geometric Approach for Efficient Consistency Assessment of Stereo Images.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Geometric consistency is crucial for accurate stereo image analysis.
• Traditional methods struggle with subtle miscalibrations in stereo pairs.
• Real-time detection of geometric distortions is needed for dynamic applications.

Purpose of the Study:

• To introduce HGC-Net, a hybrid pipeline for assessing geometric consistency in stereo image pairs.
• To develop a method capable of detecting both severe and mild geometric distortions.
• To provide an interpretable score for evaluating stereo image alignment.

Main Methods:

• Integration of classical epipolar geometry with deep learning components.
• Computation of a scalar score (A) representing the degree of geometric alignment.
• Evaluation on the Middlebury 2014 stereo dataset with synthetically distorted images.

Main Results:

• HGC-Net reliably detects subtle geometric distortions (sub-degree tilts, pixel shifts).
• The method's score degrades smoothly with increasing geometric error.
• High detection rates achieved even at minimal distortion levels, outperforming baselines.

Conclusions:

• HGC-Net offers a robust and interpretable solution for stereo image geometric consistency assessment.
• Real-time performance and independence from internal camera parameters enable broad applicability.
• Explainability features aid human operators in identifying and rectifying problematic regions.