October 26th, 2011 @ 1130am Brendan Frey (Toronto University)

Factorizing Appearance Using Epitomic Flobject Analysis

Flobject analysis uses motion or stereo disparity information to train better models of static images. During training, but not testing, optic flow is used as a cue for factorizing appearance-based image features into those belonging to different flow-defined objects, or flobjects. Here, we describe how the image epitome can be extended to model flobjects. On the tasks of classification and boundary detection, our method performs better than the original LDA-based flobject analysis technique, SIFT-based methods with and without spatial pyramid matching, and gist descriptors. Joint work with Patrick Li.

Factorizing Color and Shape Stencils Using Hierarchical Palettes

Image patches can be factorized into shape “stencils” that are shared across images and an image-specific palette that is used to paint each stencil. This leads to the notion of local palettes for patches and global palettes for entire images and ultimately to hierarchical palettes. An advantage of our approach is that color and shape can be “orthogonally” represented. Using standard classification datasets, we show that the stencil model performs better than SIFT- and GIST-based methods, even when color information is appended to their feature descriptors. Joint work with Jeroen Chua, Inmar Givoni and Ryan Adams.

seminars/seminaritems/2011-10-26.txt · Last modified: 2011/10/28 14:34 by silberman