Robin Aly defends PhD thesis on uncertainty in concept-based multimedia retrieval

by Robin Aly

This thesis considers concept-based multimedia retrieval, where documents are represented by the occurrence of concepts (also referred to as semantic concepts or high-level features). A concept can be thought of as a kind of label, which is attached to (parts of) the multimedia documents in which it occurs. Since concept-based document representations are user, language and modality independent, using them for retrieval has great potential for improving search performance. As collections quickly grow both in volume and size, manually labeling concept occurrences becomes infeasible and the so-called concept detectors are used to decide upon the occurrence of concepts in the documents automatically.

The following fundamental problems in concept-based retrieval are identified and addressed in this thesis. First, the concept detectors frequently make mistakes while detecting concepts. Second, it is difficult for users to formulate their queries since they are unfamiliar with the concept vocabulary, and setting weights for each concept requires knowledge of the collection. Third, for supporting retrieval of longer video segments, single concept occurrences are not sufficient to differentiate relevant from non-relevant documents and some notion of the importance of a concept in a segment is needed. Finally, since current detection techniques lack performance, it is important to be able to predict what search performance retrieval engines yield, if the detection performance improves.

The main contribution of this thesis is the uncertain document representation ranking framework (URR). Based on the Nobel prize winning Portfolio Selection Theory, the URR framework considers the distribution over all possible concept-based document representations of a document given the observed confidence scores of concept detectors. For a given score function, documents are ranked by the expected score plus an additional term of the variance of the score, which represents the risk attitude of the system.

User-friendly concept selection is achieved by re-using an annotated development collection. Each video shot of the development collection is transformed into a textual description which yields a collection of textual descriptions. This collection is then searched for a textual query which does not require the user's knowledge of the concept vocabulary. The ranking of the textual descriptions and the knowledge of the concept occurrences in the development collection allows a selection of useful concepts together with their weights.

The URR framework and the proposed concept selection method are used to derive a shot and a video segment retrieval framework. For shot retrieval, the probabilistic ranking framework for unobservable events is proposed. The framework re-uses the well-known probability of relevance score function from text retrieval. Because of the representation uncertainty, documents are ranked by their expected retrieval score given the confidence scores from the concept detectors.

For video segment retrieval, the uncertain concept language model is proposed for retrieving news items — a particular video segment type. A news item is modeled as a series of shots and represented by the frequency of each selected concept. Using the parallel between concept frequencies and term frequencies, a concept language model score function is derived from the language modelling framework. The concept language model score function is then used according to the URR framework and documents are ranked by the expected concept language score plus an additional term of the score's variance.

The Monte Carlo Simulation method is used to predict the behavior of current retrieval models under improved concept detector performance. First, a probabilistic model of concept detector output is defined as two Gaussian distributions, one for the shots in which the concept occurs and one for the shots in which it does not. Randomly generating concept detector scores for a collection with known concept occurrences and executing a search on the generated output estimates the expected search performance given the model's parameters. By modifying the model parameters, the detector performance can be improved and the future search performance can be predicted.

Experiments on several collections of the TRECVid evaluation benchmark showed that the URR framework often significantly improve the search performance compared to several state-of-the-art baselines. The simulation of concept detectors yields that today's video shot retrieval models will show an acceptable performance, once the detector performance is around 0.60 mean average precision. The simulation of video segment retrieval suggests, that this task is easier and will sooner be applicable to real-life applications.

[download pdf]