Alexander Haubold

Semantic concept-based query expansion and re-ranking for multimedia retrieval

We study the problem of semantic concept-based query expansion and re-ranking for multimedia retrieval. In particular, we explore the utility of a fixed lexicon of visual semantic concepts for automatic multimedia retrieval and re-ranking purposes. In this paper, we propose several new approaches for query expansion, in which textual keywords, visual examples, or initial retrieval results are analyzed to identify the most relevant visual concepts for the given query. These concepts are then used to generate additional query results and/or to re-rank an existing set of results. We develop both lexical and statistical approaches for text query expansion, as well as content-based approaches for visual query expansion. In addition, we study several other recently proposed methods for concept-based query expansion. In total, we compare 7 different approaches for expanding queries with visual semantic concepts. They are evaluated using a large video corpus and 39 concept detectors from the TRECVID-2006 video retrieval benchmark. We observe consistent improvement over the baselines for all 7 approaches, leading to an overall performance gain of 77% relative to a text retrieval baseline, and a 31% improvement relative to a state-of-the-art multimodal retrieval baseline.

Augmented segmentation and visualization for presentation videos

We investigate methods of segmenting, visualizing, and indexing presentation videos by both audio and visual data. The audio track is segmented by speaker, and augmented with key phrases which are extracted using an Automatic Speech Recognizer (ASR). The video track is segmented by visual dissimilarities and changes in speaker gesturing, and augmented by representative key frames. An interactive user interface combines a visual representation of audio, video, text, key frames, and allows the user to navigate presentation videos. User studies with 176 students of varying knowledge were conducted on 7.5 hours of student presentation video (32 presentations). Tasks included searching for various portions of presentations, both known and unknown to students, and summarizing presentations given the annotations. The results are favorable towards the video summaries and the interface, suggesting faster responses by a factor of 20% compared to having access to the actual video. Accuracy of responses remained the same on average. Follow-up surveys present a number of suggestions towards improving the interface, such as the incorporation of automatic speaker clustering and identification, and the display of an abstract topological view of the presentation. Surveys also show alternative contexts in which students would like to use the tool in the classroom environment.

Alignment of Speech to Highly Imperfect Text Transcriptions

We introduce a novel and inexpensive approach for the temporal alignment of speech to highly imperfect transcripts from automatic speech recognition (ASR). Transcripts are generated for extended lecture and presentation videos, which in some cases feature more than 30 speakers with different accents, resulting in highly varying transcription qualities. In our approach we detect a subset of phonemes in the speech track, and align them to the sequence of phonemes extracted from the transcript. We report on the results for 4 speech-transcript sets ranging from 22 to 108 minutes. The alignment performance is promising, showing a correct matching of phonemes within 10, 20, 30 second error margins for more than 60 %, 75 %, 90 % of text, respectively, on average. For perfect manually generated transcripts, more than 75 % of text is correctly aligned within 5 seconds.

Semantic Multimedia Retrieval using Lexical Query Expansion and Model-Based Reranking

We present methods for improving text search retrieval of visual multimedia content by applying a set of visual models of semantic concepts from a lexicon of concepts deemed relevant for the collection. Text search is performed via queries of words or fully qualified sentences, and results are returned in the form of ranked video clips. Our approach involves a query expansion stage, in which query terms are compared to the visual concepts for which we independently build classifier models. We leverage a synonym dictionary and WordNet similarities during expansion. Results over each query are aggregated across the expanded terms and ranked. We validate our approach on the TRECVID 2005 broadcast news data with 39 concepts specifically designed for this genre of video. We observe that concept models improve search results by nearly 50% after model-based re-ranking of text-only search. We also observe that purely model-based retrieval significantly outperforms text-based retrieval on non-named entity queries

VAST MM: multimedia browser for presentation video

In the domain of candidly captured student presentation videos, we examine and evaluate approaches for multi-modal analysis and indexing of audio and video. We apply visual segmentation techniques on unedited video to determine likely changes of topics. Speaker segmentation methods are employed to determine individual student appearances, which are linked to extracted headshots to create a visual speaker index. Videos are augmented with time-aligned filtered keywords and phrases from highly inaccurate speech transcripts. Our experimental user interface, the VAST MM Browser (Video Audio Structure Text Multi Media Browser), combines streaming videos, visual, and textual indices for browsing and searching. We evaluate the UI and methods in a large engineering design course. We report on observations and statistics collected over 4 semesters and 598 student participants. Results suggest that our video indexing and retrieval approach is effective, and that our continuous improvements are reflecting in an increase in precision and recall of user study tasks.

Web-based information content and its application to concept-based video retrieval

Semantic similarity between words or phrases is frequently used to find matching correlations between search queries and documents when straightforward matching of terms fails. This is particularly important for searching in visual databases, where pictures or video clips have been automatically tagged with a small set of semantic concepts based on analysis and classification of the visual content. Here, the textual description of documents is very limited, and semantic similarity based on WordNets cognitive synonym structure, along with information content derived from term frequencies, can help to bridge the gap between an arbitrary textual query and a limited vocabulary of visual concepts. This approach, termed concept-based retrieval, has received significant attention over the last few years, and its success is highly dependent on the quality of the similarity measure used to map textual query terms to visual concepts. In this paper, we consider some issues of semantic similarity measures based on Information Content (IC), and propose a way to improve them. In particular, we note that most IC-based similarity measures are derived from a small and relatively outdated corpus (the Brown corpus), which does not adequately capture the usage pattern of many contemporary terms: for example, out of more than 150,000 WordNet terms, only about 36,000 are represented. This shortcoming reflects very negatively on the coverage of typical search query terms. We therefore suggest using alternative IC corpora that are larger and better aligned with the usage of modern vocabulary. We experimentally derive two such corpora using the WWW Google search engine, and show that they provide better coverage of vocabulary, while showing comparable frequencies for Brown corpus terms. Finally, we evaluate the two proposed IC corpora in the context of a concept-based video retrieval application using the TRECVID 2005, 2006, and 2007 datasets, and we show that they increase average precision results by up to 200%.

Analysis and visualization of index words from audio transcripts of instructional videos

We introduce new techniques for extracting, analyzing, and visualizing textual contents from instructional videos of low production quality. Using automatic speech recognition, approximate transcripts (/spl ap/75% word error rate) are obtained from the originally highly compressed videos of university courses, each comprising between 10 to 30 lectures. Text material in the form of books or papers that accompany the course are then used to filter meaningful phrases from the seemingly incoherent transcripts. The resulting index into the transcripts is tied together and visualized in 3 experimental graphs that help in understanding the overall course structure and provide a tool for localizing certain topics for indexing. We specifically discuss a transcript index map, which graphically lays out key phrases for a course, a textbook chapter to transcript match, and finally a lecture transcript similarity graph, which clusters semantically similar lectures. We test our methods and tools on 7 full courses with 230 hours of video and 273 transcripts. We are able to extract up to 98 unique key terms for a given transcript and up to 347 unique key terms for an entire course. The accuracy of the Textbook Chapter to Transcript Match exceeds 70% on average. The methods used can be applied to genres of video in which there are recurrent thematic words (news, sports, meetings, etc.).

Classification of video events using 4-dimensional time-compressed motion features

Among the various types of semantic concepts modeled, events pose the greatest challenge in terms of computational power needed to represent the event and accuracy that can be achieved in modeling it. We introduce a novel low-level visual feature that summarizes motion in a shot. This feature leverages motion vectors from MPEG-encoded video, and aggregates local motion vectors over time in a matrix, which we refer to as a motion image. The resulting motion image is representative of the overall motion in a video shot, having compressed the temporal dimension while preserving spatial ordering. Building motion models using this feature permits us to combine the power of discriminant modeling with the dynamics of the motion in video shots that cannot be accomplished by building generative models over a time series of motion features from multiple frames in the video shot. Evaluation of models built using several motion image features in the TRECVID 2005 dataset shows that use of this novel motion feature results an average improvement in concept detection performance by 140% over existing motion features. Furthermore, experiments also reveal that when this motion feature is combined with static feature representations of a single keyframe from the shot such as color and texture features, the fused detection results in an improvement between 4 to 12% over the fusion across the static features alone.

Analysis and interface for instructional video

We present a new method for segmenting, and a new user interface for indexing and visualizing, the semantic content of extended instructional videos. Using various visual filters, key frames are first assigned a media type (board, class, computer, illustration, podium, and sheet). Key frames of media type board and sheet are then clustered based on contents via an algorithm with near-linear cost. A novel user interface, the result of two user studies, displays related topics using icons linked topologically, allowing users to quickly locate semantically related portions of the video. We analyze the accuracy of the segmentation tool on 17 instructional videos, each of which is from 75 to 150 minutes in duration (a total of 40 hours); it exceeds 96%.

Annual research review: progress in using brain morphometry as a clinical tool for diagnosing psychiatric disorders.

Brain morphometry in recent decades has increased our understanding of the neural bases of psychiatric disorders by localizing anatomical disturbances to specific nuclei and subnuclei of the brain. At least some of these disturbances precede the overt expression of clinical symptoms and possibly are endophenotypes that could be used to diagnose an individual accurately as having a specific psychiatric disorder. More accurate diagnoses could significantly reduce the emotional and financial burden of disease by aiding clinicians in implementing appropriate treatments earlier and in tailoring treatment to the individual needs. Several methods, especially those based on machine learning, have been proposed that use anatomical brain measures and gold-standard diagnoses of participants to learn decision rules that classify a person automatically as having one disorder rather than another. We review the general principles and procedures for machine learning, particularly as applied to diagnostic classification, and then review the procedures that have thus far attempted to diagnose psychiatric illnesses automatically using anatomical measures of the brain. We discuss the strengths and limitations of extant procedures and note that the sensitivity and specificity of these procedures in their most successful implementations have approximated 90%. Although these methods have not yet been applied within clinical settings, they provide strong evidence that individual patients can be diagnosed accurately using the spatial pattern of disturbances across the brain.