Cooperative Segmentation and Stereo Matching

Раздел: Статьи

Конференции] [ Книги] [ Статьи] [ Журналы] [ Все материалы] [ На главную страницу
Просмотр по: [ Годам] [ Авторам] [ Названиям]

Vinet Laurent,Sander Peter,Cohen Laurent,Gagalowicz Andre

Cooperative Segmentation and Stereo Matching

Category:
Keywords:

abstract:

Stereo matching based on points and lines as the entities has become a well developed industry. In this paper, we investigate region based matching as we feel that many of the shortcomings inherent in other approaches can be overcome by taking more developed entities (to cite but two examples: mismatches over pairs of line elements are to be expected frequently due to the lack of distinguishing features; and occlusion can effect points or segments more than regions). We present a novel approach to the combined problem of image segmentation and object matching, based on interaction between a segmentation component and a stereo component. We believe the combination to be better than the parts taken individually. This paper is limited to segmentation/matching results we are not concerned with the explicit depth computations here (but see]1]). The basic idea developed in this paper is that, since objects in the world give rise to events in both stereo images, segmentation in each image should be carried out in conjunction with segmentation in the other, thus, hopefully, producing a more reliable segmentation in both. Of course, a `vicious circle' arises in that cooperative segmentation presupposes matching, and matching is dependent on a prior segmentation. We propose breaking the circle by iteratively using partial segmentation results to suggest tentative matches, which then feed back into the segmentation procedure, and so on. Thus, segmentation (by merging and splitting regions) in one image depends on what matches have been found with the other. Some of the computation can be done independently, however, prior to any matching: a number of (candidate) segmentations of the images are computed for a range of parameters and organized in a tree structure so that merging/splitting regions just amounts to moving up/down in the tree. The complete procedure consists of two steps (see]2] for more details): 1. computing, independently for each image, to coarse hierarchical candidate segmentations; 2. determining a ` nal' segmentation from among the candidates, cooperatively with region based stereo matching between the images

На статью

Конференции] [ Книги] [ Статьи] [ Журналы] [ Все материалы] [ На главную страницу

Просмотр по: [ Годам] [ Авторам] [ Названиям]

Graphics & Media Lab. >> Библиотека | Курсы | Графикон

Верстка, дизайн, программирование Людмила Попова
Выпускающий редактор Дмитрий Ватолин
Подготовка материала Ольга Карпенко