Feature Augmentation for Learning Confidence Measure in Stereo Matching.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Confidence estimation refines stereo matching but learning-based methods often ignore spatial coherence.
• Pixel-wise confidence estimation limits performance in challenging scenarios.

Purpose of the Study:

• To develop a novel confidence estimation approach for stereo matching that incorporates spatial consistency.
• To improve the robustness and accuracy of stereo matching, particularly for complex scenes.

Main Methods:

• Extracted superpixel-level features using Gaussian mixture models and combined them with pixel-level features.
• Employed adaptive filtering in the feature domain and random regression forests for confidence map estimation.
• Utilized K-nearest neighbor aggregation for further refinement of confidence maps at both pixel and superpixel levels.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches.
• Experimental results validated the effectiveness on various benchmarks, including difficult outdoor scenes.
• The spatial consistency imposition significantly improved confidence estimation accuracy.

Conclusions:

• The novel approach effectively integrates spatial information into confidence estimation for stereo matching.
• This method offers a significant advancement for refining stereo matching results, especially in challenging visual conditions.