Enhancing the Ground Truth Disparity by MAP Estimation for Developing a Neural-Net Based Stereoscopic Camera

Abstract

This paper presents a novel method to enhance ground truth disparity maps generated by Semi-Global Matching (SGM) using Maximum a Posteriori (MAP) estimation. SGM, while not producing visually appealing outputs like neural networks, offers high disparity accuracy in valid regions and avoids the generalization issues often encountered with neural network-based disparity estimation. However, SGM struggles with occlusions and textureless areas, leading to invalid disparity values. Our approach, though relatively simple, mitigates these issues by interpolating invalid pixels using surrounding disparity information and Bayesian inference, improving both the visual quality of disparity maps and their usability for training neural network-based commercial depth-sensing devices. Experimental results validate that our enhanced disparity maps preserve SGM’s accuracy in valid regions while improving the overall performance of neural networks on both synthetic and real-world datasets. This method provides a robust framework for advanced stereoscopic camera systems, particularly in autonomous applications.

Keywords

• stereo vision
• deep learning
• disparity map
• MAP estimation
• Semi-Global Matching
• neural network