CONF odobez-CIVR-03/IDIAP Spectral Structuring of Home Videos Odobez, Jean-Marc Gatica-Perez, Daniel Guillemot, Maël EXTERNAL https://publications.idiap.ch/attachments/reports/2003/odobez_2003_civr.pdf PUBLIC https://publications.idiap.ch/index.php/publications/showcite/odobez-rr-02-55 Related documents International Conference on Image and Video Retrieval (CIVR'03) Lecture Notes in Computer Science 2003 Springer Verlag Urbana-Champaign, USA July 2003 Similar to IDIAP-RR 02-55 Accessing and organizing home videos present technical challenges due to their unrestricted content and lack of storyline. In this paper, we propose a spectral method to group video shots into scenes based on their visual similarity and temporal relations. Spectral methods have been shown to be effective in capturing perceptual organization features. In particular, we investigate the problem of automatic model selection, which is currently an open research issue for spectral methods, and propose measures to assess the validity of a grouping result. The methodology is used to group scenes from a six-hour home video database, and is assessed with respect to a ground-truth generated by multiple people. The results indicate the validity of the proposed approach, both compared to existing techniques as well as the human ground-truth. REPORT odobez-rr-02-55/IDIAP On Spectral Methods and the Structuring of Home Videos Odobez, Jean-Marc Gatica-Perez, Daniel Guillemot, Maël EXTERNAL https://publications.idiap.ch/attachments/reports/2002/rr02-55.pdf PUBLIC Idiap-RR-55-2002 2002 IDIAP Martigny, Switzerland Published in International Conference on Image and Video Retrieval, 2003 Accessing and organizing home videos present technical challenges due to their unrestricted content and lack of storyline. In this paper, we propose a spectral method to group video shots into scenes based on their visual similarity and temporal relations. Spectral methods exploit the eigenvector decomposition of a pair-wise similarity matrix and can be effective in capturing perceptual organization features. In particular, we investigate the automatic selection of the number of clusters, which is currently an open research issue for spectral methods. We first analyze the behaviour of the algorithm with respect to variations in the number of clusters, and then propose measures to assess the validity of a grouping result. The methodology is used to group scenes from a six-hour home video database, and is assessed with respect to a ground-truth generated by multiple humans. The results indicate the validity of the proposed approach, both compared to existing techniques as well as the human ground-truth.