Abstract

Video segmentation aims to consistently group regions that are similar in appearance and movement along a sequence. This task is an essential step of video analysis and it has important applications in video coding, indexing and retrieval, 3D reconstruction, action recognition, etc. In particular, in this thesis, a multiresolution hierarchical co-clustering technique (MRHC) is analyzed in order to include depth information for improving image segmentation in sequences with small variations. This information is invariant to changes on brightness and/or texture and camera position. Thus, it may correct some errors that are present when segmentation techniques based on color and movement are used, because this information is independent of color information of the image and the movement that occurs in the scene.

Two considerations can be made. Firstly, all the regions of the same object should share similar depth values and, secondly, depth values from regions that do not belong to the same object show discontinuities. Three different ways of coding depth information in MRHC have been studied in this project. In the first approach, the similarity between regions is weighted according to the depth difference between them. The second approach determines the 3D-neighborhood between regions. Finally, a combination of the previous approaches is considered.

The Video Occlusion/Object Boundary Dataset has been used to evaluate the inclusion of the depth on MRHC and to compare this method with the state-of-the-art techniques in the field of video segmentation. The results obtained show that the use of depth information improves the outcome video techniques obtained with the MRHC, outperforming the state-of-the-art methods in this scenario.