Scene Recognition and Weakly Supervised Object Localization with Deformable Part-Based Models
Megha Pandey and Svetlana Lazebnik
Dept. of Computer Science, University of North Carolina at Chapel Hill
Abstract:Weakly supervised discovery of correspondence among a set of complex, cluttered images is one of the key problems in recognition. We address this problem using deformable part-based models (DPM's) with latent SVM training. This method has originally been developed for fully supervised training of object detectors, but we demonstrate that it is also capable of more open-ended discovery of latent common structure for tasks such as scene recognition and weakly-supervised object localization. For scene recognition, DPM's can successfully capture recurring visual elements and even common objects, yielding a powerful scene representation that complements popular global image features to obtain the highest to date classification results on a challenging 67-category indoor scene dataset. For weakly supervised object localization, optimization over latent DPM parameters can successfully discover the spatial extent of objects in cluttered training images without ground-truth bounding boxes. The resulting method outperforms a recent state-of-the-art weakly supervised object localization approach on the PASCAL-07 dataset.