Research Projects
- Finding Jim Gray - Winter 2008 - Statistical Detection Theory in Psychology Seminar - Jim Gray was a famous Microsoft researcher who disappeared on his yacht sparking a internet driven search of aerial data. This project looked into how to model uncertainty in human labels of noisy imagery, applicable to active learning of machine classifiers to aid such search efforts.
- Kwote search engine - Winter 2008 - Web Search - this project extracted and attributed quotes from web news stories and presented a search application based on the open source Lucene project. Interestingly, Google announced the same feature in their news page a month after the quarter finished. When an idea is ready!
- EEG analysis - Fall 2006 - Data Mining - this project took an EEG dataset from Miguel Eckstein's group in the Psychology department and data mined using linear SVM networks to determine the ideal channels and time windows in order to discriminate which of two classes a subject was observing - face vs. car.
SpiritTagger: A Location Aware Tag Suggestion Tool Mined from Flickr
Mining online collaborative communities is a powerful emerging source for using metadata to improve existing image analysis techniques. In this work with Emily Moxley, we use a large data set of 1.5 million geo-referenced images to derive appropriate tags to suggest given the location where a photo was taken.
Spatial pyramid mining for object detection in natural images
Spatial pyramid mining is a multiresolution method for discovering commonly occurring configurations of local interest points for classes of objects in a database. The method can detect logos in a variety of realizations in real world scenes, and logos can be recognized efficiently by indexing configuration rules at multiple quantization levels.
TRECVID 2007 concept detection and search
TRECVID is an international evaluation benchmark for assessing techniques in video retrieval. The VRL at UCSB took part in concept detection, automatic detection of a set of pre-defined high level features in shots of videos, and automatic search, retrieval of video shots based on arbitrary text and image queries. We had encouraging results for a first year team in experiments fusing features based on MPEG-7 descriptors, local interest points, and audio.
TRECVID 2007 BBC rushes video summarization
In 2007 TRECVID introduced the rushes summarization pilot task. The goal was to automate the production of quality summaries no more than 4% of the original length of director b-roll films from the BBC archives. UCSB did particularly well at this task, and we were invited to talk at the ACM-MM workshop about our feature fusion method.
The multimodal music stand
UCSB hosts an IGERT in interactive digital multimedia and part of the traineeship includes working on collaborative projects. Students from ECE, Music, and Media Arts & Technology department worked in JoAnn Kuchera-Morin's group to develop a new musical interface - a multimodal music stand that can perform gesture recognition and also operate itself as a theremin-based musical instrument.