Effective Part Localization in Latent-SVM Training

Deformable part models show a remarkable detection performance for a variety of object categories. During training these models rely on energy-based methods and heuristic initialization to search and localize parts, equivalent to learning local object features. Due to weak supervision, however, those learnt part detectors contain lots of noise and are not enough reliable to classify the object. This paper investigates part localization problem and extends the latent-SVM by incorporating local consistency of image features. The objective is to adaptively select part sub-windows that overlap semantically meaningful components as much as possible, which leads to a more reliable learning of the part detectors in a weakly-supervised setting. The main idea of our method is estimating part-specific color/texture models as well as edge distribution within each training example, followed by a foreground segmentation for part localization. The experimental results show that we achieve an overall improvement of about 3% mAP over the latent-SVM on non-rigid objects.