英文摘要 | 3D reconstruction, i.e. to recover scenes’ visible surfaces from images, has been the fundamental problem in computer vision. Recently, reconstructions are increasingly used on applications such as digital earth, digital archaeology and digital entertainment, where the visual quality of reconstructed scenes rather than the strict shape, becomes the major concern. The high quality visual effect needs dense or at least quasi-dense matches, who cannot be obtained by currently popular feature extracting and matching approaches, for they can only find sparse correspondences, and sparse correspondences are more suitable for estimating epipolar geometry and camera parameters. Dense stereo matching has been studied more thoroughly and can produce promising results, but is much more costly at the same time. Moreover, stereo matching methods often require strict camera configuration. On contrast, match propagation, a promising strategy, can make a good balance between these two aspects. In this dissertation, the main work is focused on how to find quasi-dense matches through match propagation, i.e. how to propagate matches from some sparse matches to their neighborhoods. The main work is 3-fold: · In order to meet the needs of reconstruction, an improved match propagation algorithm for perspective images is proposed, implemented and tested. The issues of getting precise seed matches and propagation strategy are thoroughly discussed. The algorithm can produce not only well-distributed but also accurate matches. In addition, some practical issues in real applications are also discussed. · A novel quasi-dense matching algorithm through local neighborhood transfer is proposed for fish-eye images, characterized by severe projective distortions. A local affine model is used for the geometric transformation between two corresponding image patches. Patches are normalized with the known affine transformation before match propagation is applied to them. The local affine transformation is computed during the propagation by a new mechanism, called neighborhood-transfer, to ensure the propagation be carried out in the correct neighborhood. Experiments with real fish-eye images show that our algorithm can provide satisfactory results. · A unified framework is proposed which could accommodate match propagation between different kinds of images, including perspective images, fish-eye images and catadioptric images. This unified framework is applicable for both po... |
修改评论