Active Learning for Information Retrieval : Using 3 D Models As An Example

In this paper, we propose a general approach to content-based information retrieval. We use active learning to produce hidden annotations to improve the retrieval performance. We apply the proposed algorithm to designing a 3D model retrieval system. Our database is composed of approximately 1750 3D models having among 53 attributes that form a complex attribute tree. For each model in the database, we maintain a list of probabilities, each indicating the probability of this model having one of the attributes. During training, the learning algorithm samples models in the database and presents them to the annotator to assign attributes to. For each sampled model, each probability is set to be one or zero depending on whether or not the corresponding attribute is assigned by the annotator. For models that have not been annotated, the learning algorithm estimates their probabilities based on a potential function. Specifically, models close to an annotated model are likely to have the same attributes as the annotated model, so similar probabilities from the annotated model are given to these models. At each stage of the learning, the algorithm is able to determine, among the models that have not been annotated, which model the system is the most uncertain of, and presents it as the next sample to the annotator to assign attributes to. During retrieval, the list of probabilities works as a feature vector for us to calculate the semantic distance between two models, or between the user query and a model in the database. The overall distance between two models is determined by a weighted sum of the semantic distance and the low-level feature distance. We show that with the proposed algorithm, the retrieval performance of the system improves rapidly with the number of annotated samples. Furthermore, we show that active learning outperforms learning based on random sampling.