StereoSGBM

StereoSGBM, which stands for StereoSonographic Gradient-based Matching, is a stereo vision algorithm used for computing dense disparity maps. It is a variant of the Semi-Global Matching (SGM) algorithm, designed to be more computationally efficient and to produce smoother disparity maps.

The core idea behind StereoSGBM is to combine the strengths of local and global stereo matching techniques. It performs a pixel-wise matching cost calculation, similar to local methods, but then incorporates a smoothness constraint that propagates information across the image. This is achieved by aggregating costs along multiple paths in the image, effectively enforcing that neighboring pixels with similar colors should have similar disparities.

Unlike traditional SGM which aggregates costs over many paths, StereoSGBM typically uses a smaller, fixed set of paths, often four or eight, to reduce computational complexity. This trade-off allows for faster processing while still achieving a significant improvement in disparity map quality over purely local methods.

The algorithm starts by computing a matching cost volume, where each entry represents the dissimilarity between a pixel in the left image and a pixel in the right image at a specific disparity. This cost is often based on pixel intensity differences or color information. Then, the algorithm aggregates these costs along the chosen paths, penalizing large disparity changes between adjacent pixels. Finally, the disparity for each pixel is determined by finding the disparity that minimizes the aggregated cost. StereoSGBM is widely used in applications requiring real-time or near real-time depth estimation, such as autonomous driving, robotics, and augmented reality.