[Go] wide baseline and narrow baseline

The purpose of the baseline is to refer to the distance between the two camera light cores in a stereo vision system. The corresponding point matching problem can be divided into a wide baseline (Wide Baseline) and a narrow baseline match (short Baseline) based on the viewpoint position relationship of the two images taken. When the term "wide baseline" is used for matching, it generally refers to the matching of two images with obvious differences. This can be caused by a large difference in position between the cameras and possibly due to factors such as camera rotation or changes in focal length.

The demarcation between broad baseline and narrow baseline matching is not very strict, but in narrow baseline matching there are the following assumptions: camera focal length and other internal parameters vary little: The camera position will not be far apart, there will be no large rotation, the corresponding point of the neighborhood is similar.

The following hypothesis exists in the wide baseline match: The corresponding point on the image can be any position, the camera can move arbitrarily, and the focal length and other parameters of the camera can be changed greatly; the scene on one image may be obscured on the other; the neighboring region of the corresponding point has a similar place, However, due to the change of camera position and illumination, the similarity of single dependent neighborhood cannot be correctly corresponded.

The typical method in narrow baseline matching is to utilize the neighbor cross-correlation (neighborhood cross-correlation) method. But in the case of wide baseline, the distance between images is far, the imaging conditions are very different, even if the same characteristics of space, in the image of the optical characteristics (gray value, color value, etc.), geometric characteristics (shape, size, etc.) and spatial location (image position, direction, etc.) are very big difference, plus noise, Occlusion and other factors exist, the matching method based on neighborhood correlation is invalid. In the wide baseline matching, it is difficult to match the information of the characteristic itself, such as the position information of the edge and corner, and the researchers combine several features, especially the structural features, to form a stable eigenvector (called the feature descriptor). This characteristic vector, which maintains certain stability to the geometric deformation and illumination variation of the image, is called invariant (INVARIARLT). Invariant technology is an important technique in wide-baseline matching applications.

Selected from Sean Master's thesis.

Reference: Http://blog.163.com/[email protected]/blog/static/83046444201071731444339/

[Go] wide baseline and narrow baseline