M. Pawan Kumar - Self-Paced Learning for Specific-Class Semantic Segmentation

Опубликовано 18 августа 2016, 4:51

Algorithms for learning the parameters of latent variable models are prone to getting stuck in a bad local optimum. To alleviate this problem, we build on the intuition that the algorithm should be presented with the training data in a meaningful order: easy samples first, difficult samples later. As we are often not provided with a readily computable measure of easiness, we design a novel self-paced learning algorithm that simultaneously selects easy samples and learns the parameters. We empirically demonstrate that self-paced learning outperforms the state of the art on several standard applications. Next, we consider the task of learning a model that provides a complete segmentation of an image by assigning each of its pixels to a specific semantic class. The main problem we face is that lack of fully supervised data. To address this issue, we develop a principled framework for learning the parameters of a specific-class segmentation model using diverse data (with varying levels of supervision). More precisely, we formulate our problem as a latent structural support vector machine (LSVM), where the latent variables model any missing information in the human annotation. In order to deal with the noise inherent in weakly supervised annotations, we train the LSVM with self-paced learning. Using large, publicly available datasets we show that our approach is able to exploit the information offered by different annotations to improve the accuracy of specific-class segmentation.