Where: TEMICS / EQUIPE / STAUDER /

Submitted Publications Jürgen Stauder

Journal Papers

[Stau00c]: J. Stauder: "An illumination effect descriptor for video sequences", submitted to the Journal of Visual Communication and Image Representation, October 1999.; Abstract: In the context of indexing of video data bases, the future standard MPEG-7 will provide descriptors for motion, shape, texture and color to identify video scenes. In this paper, an illumination effect descriptor is presented that has been proposed to MPEG-7. Effects caused by the the scene illumination can be a useful feature for video indexing. The proposed descriptor addresses temporal changes of object shading, of cast shadows and of the global illumination intensity by one scalar. It is based on temporal changes of image luminance along motion trajectories. Further, an automatic, noise adaptive method for extraction of the illumination effect descriptor from a video sequence is presented. A performance analysis and sample retrieval experiments show that extracted descriptor values are sufficiently precise and sensible to distinguish video sequences with different scene illumination.
[Stau00b]: J. Stauder: "Object rotation axis from shading", submitted to the IEEE Transactions on Pattern Analysis and Machine Intelligence", October 1999.; Abstract: In this paper, an estimator is developed for the tilt angle of the rotation axis of a rigid object that is illuminated by a point light source. The tilt angle defines the orientation of the 2D projection of the 3D object unit rotation axis in the image plane. The angle is one of the 6 degrees of freedom of 3D rigid motion. It is analytically shown that the tilt angle can be estimated just by exploiting the error of motion compensation between two images due to object shading. Required input data are a dense displacement vector field (DVF) for motion compensation and a 2D binary object mask. The estimator assumes rigid, Lambertian objects with equally distributed surface normals as well as a scene illumination by a distant point light source and ambient light. Experiments with head-and-shoulder video sequences yield qualitatively good results that are derived fully automatically. A performance analysis shows that the estimator works best if light comes from viewing direction. Finally, an application to video indexing shows that head-and-shoulder sequences can be fully automatically segmented into temporal segments of homogeneous object motion. A demo can be found at http://www.irisa.fr/temics/Demos.

Conference Papers

[ECCV-2000b]: J. Stauder: "Performance analysis of point light source estimation", submitted to ECCV-2000, 6th European Conference on Computer Vision, 26.7.-1.7.2000, Dublin, Ireland.; Abstract: For the estimation of point light source parameters from video sequences, methods have been published recently that are fully automatic. They are not constrained to scenes with unicolored objects as e.g. the well known pioneering algorithm of Pentland, 1982. The estimated parameters are the intensity and the direction of a single distant point light source in presence of ambient light. These recent methods exploit two video images. They assume at least one moving object and need information on the 3D object motion and 3D object shape. In this paper, the performance of this family of point light source parameter estimation methods is analyzed in terms of estimation theory. Therefore, all input data and its inherent errors are described by a stochastic observation model. Based on this model, the performance is analyzed regarding the Cramér-Rao theoretical lower bound of estimation error variances. The bound is derived for a variety of cases of scene illumination, object motion and errors in input data. The analysis shows in which cases which estimation accuracy can be expected with current methods. Finally, a comparison of the bound with one of the recent estimators give insights how to improve illumination estimation.
[ECCV-2000a]: J. Stauder: "A video descriptor based on illumination effects", submitted to ECCV-2000, 6th European Conference on Computer Vision, 26.7.-1.7.2000, Dublin, Ireland.; Abstract: In the context of indexing of video data bases, the future \mbox{standard} MPEG-7 will provide descriptors for motion, shape, texture and color to identify video scenes. In this paper, an illumination effect descriptor is presented that has been proposed to MPEG-7. Effects caused by the scene illumination can be a useful feature for video indexing. The proposed descriptor addresses temporal changes of object shading, of cast shadows and of the global illumination intensity by one scalar. It is based on temporal changes of image luminance along motion trajectories. The descriptor allows for automatic, noise adaptive extraction from a video sequence. Sample retrieval experiments show that extracted descriptor values are sufficiently precise and sensible to distinguish video sequences with different scene illumination.

Webmaster:
Last time modified: 2006-03-31