Stereo Vision Techniques: A Comparative Study of Traditional and Machine Learning-Based Approaches

Stereo vision, which derives 3D information from pairs of 2D images, remains pivotal for generating precise 3D representations, although it faces challenges such as camera calibration, image rectification, and disparity map computation. In this paper, we carry out a comprehensive analysis of both traditional stereo vision algorithms (SGBM and WLS-SGBM) and advanced machine learning-based techniques (RAFT-Stereo, Selective-IGEV) for producing disparity maps and 3D point clouds. We employ both dense and non-dense strategies—generating keypoint-based and region-of-interest (ROI) point clouds—to enable consistent comparisons. Experimental results show that RAFT-Stereo and Selective-IGEV excel in producing visually coherent 3D scenes, while SGBM (enhanced with WLS) maintains uniform precision. Additionally, a hardware-based depth acquisition approach complements these algorithmic methods, and the YOLOv11-based pipeline proves valuable for targeted depth and height estimations. Overall, we demonstrate that advanced stereo vision approaches can match—and often exceed—the performance of classical block-matching algorithms, offering robust solutions for a wide range of 3D reconstruction tasks.