Final Report Summary - VIAMASS (Visual Recognition Made Super-Scalable)
The project Viamass has proposed some novel methodologies to address the problem of collection representations and similarity search in high-dimensional spaces. We focused on large image collections represented by high-dimensional vectors like those based on aggregated local descriptors or those derived from deep learning.
In this context, we have proposed several unconventional ways to address these problems, such as considering this setup as a regression problem. This led us to find a set representation optimizing the quality of the membership test. Second, we have introduced group testing for similarity search and achieved excellent results for high-dimensional vectors. Finally, we show that it is possible to produce a vector set representation that also includes some partial geometrical information, which is useful either to construct new local descriptors that automatically find the best relative angle or 2) to produce image-level representation from small image regions that aggregate local image descriptors in a way that preserves the relative orientation of the regions.
* Please, notice that this summary may be published
In this context, we have proposed several unconventional ways to address these problems, such as considering this setup as a regression problem. This led us to find a set representation optimizing the quality of the membership test. Second, we have introduced group testing for similarity search and achieved excellent results for high-dimensional vectors. Finally, we show that it is possible to produce a vector set representation that also includes some partial geometrical information, which is useful either to construct new local descriptors that automatically find the best relative angle or 2) to produce image-level representation from small image regions that aggregate local image descriptors in a way that preserves the relative orientation of the regions.
* Please, notice that this summary may be published