| 研究実績の概要 |
Although much research effort has been put into face recognition, the problem remains very difficult. The aim of this project is to use the redundancy in the database to expand the annotation of face images, by building a reasoning system with the aim of propagating as much as possible the face labels.
The outcomes of the first year of the project led us to believe that the performance of our label propagation methods depended crucially on the quality of the similarity graph used. Accordingly, in the final year of the project, we focused on what turned out to be the central issue: the dynamic modification of the similarity graph in order to enhance the quality of image labeling produced by the propagation process. The main outcome was a very successful graph construction method based on the following strategy. For each instance of a tagged image, the relevance of each of its features is assessed according to the discriminability of the similarity measure over all other tagged images, with respect to the single feature. Then, the tagged image is associated with a reduced set of its own most discriminative features. These features are then used to customize the similarity values between tagged and untagged faces in the graph construction.
These methods required the efficient on-demand computation of similarity values using subsets of the original feature set. To support such computation, we investigated fast approximate similarity search using precomputed approximations to a target similarity measure supplied at query time.
|
