The PIs have recently developed a discriminatively trained deformable part based model for detecting objects in images. This model achieves a two-fold improvement in average precision over the best performance in the 2006 PASCAL person detection challenge. Our entry in the 2007 object detection challenge won in six classes. Modifications since the 2007 challenge have improved our results so that we now outperform the best 2007 challenge results in ten of twenty classes including the person class. These accomplishments rely on a newly developed approach to discriminative training for latent variable models which we call a latent SVM (LSVM). We propose to extend this methodology in a variety of ways. The research will focus on deeper latent information such as subclassification (mixture models), three dimensional pose, and figure/ground segmentations. They will also use class hierarchies, visual words, and hierarchical object models with parts and sub-parts. We also propose a general methodology for using SVM training to train models, such as geometry-based 3D models, which are highly nonlinear in model parameters. All aspects of this research are strongly tied to empirical performance -- no method will be adopted unless it actually improves the state of the art. The goal has been, and will continue to be, to improve the state of the art through the use of semantically deeper models and improved general purpose machine learning methods.
Progress on this project can be found at http:/ttic.uchicago.edu/ ~dmcallester/objects.html
