Exploiting Competition Relationship for Robust Visual Recognition

Abstract

Leveraging task relatedness has been proven to be beneficial in many machine learning tasks. Extensive researches has been done to exploit task relatedness in various forms. A common assumption for the tasks is that they are intrinsically similar to each other. Based on this assumption, joint learning algorithms are usually implemented via some forms of information sharing. Various forms of information sharing have been proposed, such as shared hidden units of neural networks, common prior distribution in hierarchical Bayesian model, shared weak learners of a boosting classifier, distance metrics and a shared low rank structure for multiple tasks. However, another very common and important task relationship, i.e., task competition, has been largely overlooked. Task competition means that tasks are competing with each other if there are competitions or conflicts between their goals. Considering that tasks with competition relationship are universal, this dissertation is to accommodate this intuition from an algorithmic perspectives and apply the algorithms to various visual recognition problems. Focus on exploiting the task competition relationships in visual recognition, the dissertation presents three types of algorithms and applied them to different visual recognition tasks. First, hypothesis competition has been exploited in a boosting framework. The proposed algorithm CompBoost jointly model the target and auxiliary tasks with a generalized additive regression model regularized by competition constraints. This model treats the feature selection as the weak learner (\ie, base functions) selection problem, and thus provides a mechanism to improve feature filtering guided by task competition. More specifically, following a stepwise optimization scheme, we iteratively add a new weak learner that balances between the gain for the target task and the inhibition on the auxiliary ones. We call the proposed algorithm CompBoost, since it shares similar structures with the popular AdaBoost algorithm. In this dissertation, we use two test beds for evaluation of CompBoost: (1) content-independent writer identification by exploiting competing tasks of handwriting recognition, and (2) actor-independent facial expression recognition by exploiting competing tasks of face recognition. In the experiments for both applications, the approach demonstrates promising performance gains by exploiting the between-task competition relationship. Second, feature competition has been instantiated through an alternating coordinate gradient algorithm. Sharing the same feature pool, two tasks are modeled together in a joint loss framework, with feature interaction encouraged via an orthogonal regularization over feature importance vectors. Then, an alternating greedy coordinate descent learning algorithm (AGCD) is derived to estimate the model. The algorithm effectively excludes distracting features in a fine-grained level for improving face verification. In other words, the proposed algorithm does not forbid feature sharing between competing tasks in a macro level; it instead selectively inhibits distracting features while preserving discriminative ones. For evaluation, the proposed algorithm is applied to two widely tested face-aging benchmark datasets: FG-Net and MORPH. On both datasets, our algorithm achieves very promising performances and outperforms all previously reported results. These experiments, together with detailed experimental analysis, show clearly the benefit of coordinating conflicting tasks for improving visual recognition. Third, two ad-hoc feature competition algorithms have been proposed to apply to visual privacy protection problems. Visual privacy protection problem is a practical case of competition factors in real world application. Algorithms are specially designed to achieve best balance between competing factors in visual privacy protection based on different modeling frameworks. Two algorithms are developed to apply to two applications, license plate de-identification and face de-identification.