Model Recommendation for Action Recognition and Other Applications

Abstract

The typical approach to learning based vision has been that for each individual application, classifiers or detectors are learned anew from annotated training data for each specific task. However, the classifiers trained in this way tend to be brittle and highly specialized to the datasets from which they are derived, making them difficult to transfer between tasks. While multi-task learning and domain adaption techniques address some of these problems on a theoretical level, from a practical standpoint they are just as complicated and labor-intensive as the simpler learning techniques they supplant. However, suppose that these specialized classifiers had simply been collected into a library: while it is unlikely that any specific classifier would generalize well to a new dataset, there may exist some classifier in the library that is tuned to the same conditions as the new task. This thesis addresses the fundamental question of how to efficiently select a good classifier from such a library. Specifically, this thesis demonstrates that collaborative filtering techniques (such as employed by recommender systems like Netflix and Amazon.com) can be used to recommend models appropriate for a specific target task. These recommendations are made by trying, or rating, a small subset of models on the target task, and then using that small set of ratings along with the ratings of the models on other tasks to predict the ratings of the unevaluated models on the target task. This process, which we term ``model recommendation'', is applied to action recognition and other vision and robotics applications, and the subtle differences between model recommendation and typical recommender systems are used to derive novel algorithms and extensions to the core recommendation concept.