Unified Confidence Estimation Networks for Robust Stereo Matching

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Stereo matching algorithms require accurate stereo confidence for improved performance.
• Existing deep convolutional neural networks (CNNs) methods often rely on single inputs, limiting accuracy.

Purpose of the Study:

• To develop a deep architecture for estimating stereo confidence by simultaneously utilizing heterogeneous inputs.
• To enhance the accuracy of stereo matching algorithms through improved confidence estimation.

Main Methods:

• A novel deep architecture is proposed, processing both matching cost volumes and disparity maps.
• Matching probability volumes are computed using residual networks and pooling modules.
• A unified deep network extracts confidence features from probability volumes and disparity maps, employing multi-sized convolutional filters.

Main Results:

• The proposed method achieves superior stereo confidence estimation compared to state-of-the-art techniques.
• Experimental results validate the effectiveness of the approach across various benchmarks.
• The semi-supervised learning approach utilizes a novel loss function with confident points for image reconstruction.

Conclusions:

• The presented deep architecture effectively estimates stereo confidence by integrating multiple data sources.
• This method offers a significant advancement in stereo matching accuracy and robustness.
• The proposed techniques are validated through rigorous experimentation and comparison with existing methods.