Simultaneous Stereo Matching and Confidence Estimation Network.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep stereo matching is crucial for 3D reconstruction and depth perception.
• Accurate disparity and confidence estimation are vital for reliable distance imaging.
• Sequential training of disparity and confidence models can be suboptimal.

Purpose of the Study:

• To present a novel multi-task model for simultaneous disparity and confidence prediction in deep stereo matching.
• To demonstrate the advantages of parallel training over sequential training for stereo matching tasks.
• To investigate the impact of loss function weighting on model performance.

Main Methods:

• A multi-task deep learning model was developed by combining successful single-task models.
• A novel loss function was proposed for joint training of disparity and confidence prediction.
• The performance of the parallel multi-task model was compared against sequential training approaches.
• The effect of weighting loss function components was analyzed.

Main Results:

• The multi-task model achieved a 15% to 30% improvement in the Area Under the Curve (AUC) metric when trained in parallel versus sequentially.
• Investigated the impact of weighting loss function components on stereo and confidence prediction performance.
• Demonstrated that improved confidence estimation enhances the practicality of stereo estimators for generating distance images.

Conclusions:

• Simultaneous prediction using a multi-task model offers significant advantages over sequential methods in deep stereo matching.
• The proposed approach enhances the accuracy and reliability of depth estimation for practical applications.
• Optimizing confidence estimation is key to improving the utility of stereo vision systems for distance imaging.