Discriminatively Trained Mixtures of Deformable Part Models

We have developed a new system building on our work on discriminatively trained, multiscale, deformable part models [1]. As in our previous system the models are trained using a discriminative procedure that only requires bounding box labels for positive examples. Our new system uses mixture models. Each mixture component is similar to a model from [1], consisting of a coarse “root” filter and 6 part models. Each part model consists of a spatial term and a part filter. The spatial term specifies an ideal location for a part relative to the root and a quadratic deformation cost for placing the part at some other location. The score of a component in a detection window is the score of the root filter on the window plus the sum over parts, of the maximum over placements of that part, of the part filter score on its subwindow minus the deformation cost of the placement. Both root and part filters are scored by computing the dot product between a set of weights and histogram of gradient (HOG) features within a window. As in [1] the features for the part filters are computed at twice the spatial resolution of the root filter. The score of a mixture model in a detection window is the maximum score over its components, where the scores are calibrated by a component specific offset parameter. Models are defined at a fixed scale, and we detect large objects by searching over an image pyramid.