Semi-supervised and unsupervised extensions to maximum-margin structured prediction

Publication Type:
Issue Date:
Full metadata record
Files in This Item:
Filename Description Size
01front.pdf111.49 kB
Adobe PDF
02whole.pdf1.6 MB
Adobe PDF
Structured prediction is the backbone of various computer vision and machine learning applications. Inspired by the success of maximum-margin classifiers in the recent years; in this thesis, we will present novel semi-supervised and unsupervised extensions to structured prediction via maximum-margin classifiers. For semi-supervised structured prediction, we have tackled the problem of recognizing actions from single images. Action recognition from a single image is an important task for applications such as image annotation, robotic navigation, video surveillance and several others. We propose approaching action recognition by first partitioning the entire image into “superpixels”, and then using their latent classes as attributes of the action. The action class is predicted based on a graphical model composed of measurements from each superpixel and a fully-connected graph of superpixel classes. The model is learned using a latent structural SVM approach, and an efficient, greedy algorithm is proposed to provide inference over the graph. Differently from most existing methods, the proposed approach does not require annotation of the actor (usually provided as a bounding box). For the unsupervised extension of structured prediction, we considered the case of labeling binary sequences. This case is important in a detection scenario, where one is interested in detecting an action or an event. In particular, we address the unsupervised SVM relaxation recently proposed in (Li et al. 2013) and extend it for structured prediction by merging it with structural SVM. The main contribution of the proposed extension (named Well-SSVM) is a re-organization of the feature map and loss function of structural SVM that permits finding the violating labelings required by the relaxation. Experiments on synthetic and real datasets in a fully unsupervised setting reveal a competitive performance as opposed to other unsupervised algorithms such as k-means and latent structural SVM. Finally, we approached the problem of unsupervised structured prediction by M³ Networks. M³ Networks are an alternative formulation of maximum-margin structured prediction that can satisfy the complete set of constraints for decomposable feature and loss functions; hence, the entire set of constraints is considered during the search for the optimal margin as opposed to Structural SVM. In the thesis, we present the interpretation of M³ Networks in Well-SSVM, thus allowing us to use in a semi-supervised and unsupervised scenario.
Please use this identifier to cite or link to this item: