Talk by Brendan Frey - David Lowe

Speaker: Brendan Frey, University of Toronto Web page: http://psi.toronto.edu/

Brendan Frey will be giving the following two short talks, with a combined length of under an hour.   He will be visiting us for just a few hours on Thursday afternoon, so if anyone would like to meet with him, David Lowe is arranging his schedule.

Factorizing Appearance Using Epitomic Flobject Analysis

Previously, flobject analysis was introduced as a method for using motion or stereo disparity information to train better models of static images.

During training, but not during 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 and introduce a suitable learning algorithm. Using the CityCars and CityPedestrians datasets, we study the tasks of object classification and localization. Our method performs significantly better than the original LDA-based flobject analysis technique, SIFT-based methods with and without spatial pyramid matching, and gist descriptors.

Learning Structural Element Patch Models With Hierarchical Palettes

 Image patches can be factorized into shapelets that describe segmentation patterns called structural elements (stels), and palettes that describe how to paint the shapelets. We introduce local palettes for patches, global palettes for entire images and universal palettes for image collections. Using a learned shapelet library, patches from a test image can be analyzed using a variational technique to produce an image descriptor that represents local shapes and colors separately. We show that the shapelet model performs better than SIFT, Gist and the standard stel method on Caltech28 and is very competitive with other methods on Caltech101.