Objectives: Design, development and evaluation of positive identity of objects in visual search engines.
This WP builds on the extensive success of visual recognition in computer vision. What is missing in the current search engines is the evidence that the object as indicated is really there. Also, nearly identical objects or object classes cannot be discriminated. These are important cases we wish to discriminate.
The state of the art in content-driven image search is best characterized by Mean Average Precision. They sort all data along the given concept classifier outcome and present to the user the top 10 best ones, always no matter how poor the result is. There is no notion whether the object or scene is actually found. To find a specific instance, a redesign of search engines is needed: positive identification resolution. They will require
- new features,
- positive object localization,
- machine learning at a different operating point on the ROC curve.
The positive search engine would find fewer examples, and make more sense in professional use. This general challenge generates two sub-challenges:
- Expanding visual feature sets into hierarchically ordered representations for scenes and objects separately based on decomposable hierarchical orderings.
- Precision is enhanced by applying a tailored sequence of verifiers for each query.