posted by Richard Power
The year 2009 dawned in anxiety, as financial institutions throughout the world shook to their foundations, as if from an earthquake. But this earthquake did not issue from the shifting of tectonic, but from inside the financial system itself; and as the reverberations spread outward from the epicenter, a tsunami swept through the global economy.
In such circumstances, the future cannot arrive soon enough; and the future, after all, is the business of CyLab. CyLab faculty and graduate students are working on seven research thrusts, and along seven more cross-cutting research thrusts, in an audacious program aimed at harnessing the future to secure the present; and, of course, in the process, they are contributing to renewed prosperity and opportunity through capacity building in the areas of technology, personnel and industry.
In this nineteenth issue of CyLab Chronicles, we will explore the dynamic matrix of CyLab from two different perspectives: first, from the ten thousand foot level of the research thrusts, and then from the facts-on-the-ground of CyLab’s weekly webinar series. (NOTE: Access to the weekly webinar series is a benefit of the CyLab partnership program.)
The full scope of the CyLab research program, with its seven thrusts and seven additional cross-cutting thrusts, is depicted in the accompanying diagram.
These thrusts, in turn, resolve upward into the six broader, overarching elements, recently articulated by CyLab faculty:
Taking on some of the most challenging issues in the field, the CyLab Biometrics Lab has become a force to reckon with. In his presentation on Super-Resolution for Face Recognition for the CyLab weekly webinar series, Vijayakumar Bhagavatula offered a compelling example of how and why.
Here is a brief glimpse into Professor Bhagavatula’s talk.
“This is the work of Pablo Hennings, who is sitting here in the front row to keep me honest, it was his PhD. thesis work; I just get to have the fun of presenting it. We also had a lot of help from Dr. Simon Baker, who used to be in the Robotics Institute here at Carnegie Mellon, but now is at Microsoft Research.”
In the areas of both surveillance and forensics, there is an urgent need for super-resolution to add in the effort to optimize face recognition processes, e.g.:
There are other potential uses less related to cloak and dagger, e.g.:
But none of the reasons is more poignant than the potential uses in the effort to find missing persons, in particular missing children:
With super-resolution face recognition capabilities, surveillance camera networks already in place to perform other tasks could be utilized to deepen and intensify the search for these missing children.
“There are video cameras all over the world now -- many intersections, many shopping malls. If only we could use the surveillance videos produced by those cameras and be able to match what is in those surveillance videos with these reference images we have of these missing people.”
“Of course, resolution is only one of the issues. There are a lot of other issues – lighting levels, expression, pose, etc. Face recognition has a lot of challenges. My talk today is only focused on the resolution issue. I do not want to paint the picture that this will solve everything. This addresses one issue.”
“The punch line is that most super-resolution methods that are out there are aimed at reconstruction, creating good looking images for human consumption. But good looking images are not necessarily easy to recognize face images. The key is to recognize, i.e., face recognition. So how do we do super-resolution with the aim that we want to recognize these faces, not just reconstruct them?
“There are two ways to address this problem, in which we have high resolution training images and low resolution test image. There are two obvious approaches. One is to take the low resolution test images, apply a super-resolution algorithm to get higher resolution images, and then do the matching there. That produces artifacts, etc. The other approach is to go from high resolution gallery, or training images, and bring them down to the low resolution that your test images are going to be, and do your matching in the low resolution. And there you are giving up information, obviously.”
“But there is another way to look at it, this is Pablo’s contribution, and we call it S2R2, Simultaneous Super-Resolution Recognition. With S2R2, you do not use either of the two methods I just mentioned, but to actually use the high resolution gallery and the low resolution test images simultaneously to do recognition.”
We have created several publications to help tell the story of CyLab, as it pursues its mission of harnessing the future to secure the present:
See all CyLab Chronicles articles