Uncertainty guided ranking loss for enhanced domain generalizable stereo-matching

• Main Author: Nallapati, Nikhil
• Other Authors: Lam Siew Kei
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/180792
• Institution: Nanyang Technological University

Stereo-matching is a fundamental technique for accurately estimating scene depth in numerous computer vision applications. However, prevailing deep learning-based stereo-matching networks frequently encounter challenges regarding generalization across different domains. Recent research endeavors have addressed this issue primarily by mitigating the detrimental impact of shortcut features stemming from excessive dependence on local chromatic attributes such as color, illumination, and texture. In this thesis, we propose a novel approach to enhance the generalization capabilities of stereo-matching networks from a geometric standpoint. Specifically, we introduce a geometry regularization technique that encodes depth ordinal relationships through a stereo uncertainty-guided ranking loss mechanism. This regularization strategy aims to guide the network towards effectively navigating challenging areas during the learning process, ultimately facilitating a more accurate geometrical understanding of the scene. Our proposed regularization scheme complements existing research efforts to improve the generalization of stereo-matching networks at the chromaticity level. Furthermore, it seamlessly integrates with these frameworks, resulting in additional performance enhancements. Through extensive evaluations conducted on diverse datasets, we demonstrate the effectiveness of our method in achieving superior performance. To ascertain the practical viability of our methodology, a field trial was conducted leveraging the ZED 2i stereo camera in conjunction with a real-time model deployed on an Nvidia Jetson development kit. The efficacy of our approach was thoroughly demonstrated through these experimental trials. Comprehensive evaluations were conducted using real-world images captured within the NTU campus environment, thereby substantiating the effectiveness and applicability of the proposed methodology. Our thesis contributes to advancing the field of stereo-matching by proposing a novel geometry regularization approach that enhances the generalization capabilities of stereo-matching networks, leading to improved depth estimation accuracy across various scenarios and environments.