Colloquium: Seeing Camouflaged Objects and Unsupervised Joint Alignment

Computer Science Department Colloquium

Title: Seeing Camouflaged Objects and Unsupervised Joint Alignment

Speaker: Professor Erik Learned-Miller
         Department of Computer Science
         University of Massachusetts Amherst

Location: Marksbury Theater 
Time: 4pm, Monday Apr 20, 2015
Host: Professor Nathan Jacobs


Part I: Seeing camouflaged objects.

This part of the talk is about detecting moving objects.The ultimate challenge for these methods is to see well-camouflaged objects, such as animals with natural camouflage, which are virtually undetectable until they move. To make the problem even more challenging, we consider the detection of camouflaged objects when the camera is moving relative to the background. We detect these objects by segmenting them based upon their motion. In particular, we estimate the parameters of the background motion (translation and rotation) and highlight objects that are moving in a way that is inconsistent with this motion. We demonstrate our methods on some challenging videos including videos of naturally camouflaged animals.

Part II: Unsupervised joint alignment.

Aligning images to a canonical position can aid significantly in recognizing those images. For example, in face recognition, orienting the face so that it is well-centered in the image can significantly improve face recognition algorithm accuracy. I review our previous work in unsupervised joint alignment, known as "congealing" and "funneling". I describe a new project in which we are attempting to extend this technique to do unsupervised pose estimation of faces. That is, we are developing algorithms that can estimate the pose of a head relative to other poses without ever being given training examples of those poses.

Speaker Bio: Ph.D., Electrical Engineering and Computer Science, Massachusetts Institute of Technology (2002), M.S., Electrical Engineering and Computer Science, Massachusetts Institute of Technology (1997), B.A., Psychology, Yale University (1988). Professor Learned-Miller joined the faculty in the University of Massachusetts Amherst Department of Computer Science in 2004 as an Assistant Professor. His most recent position was a post-doctoral research engineer in the Electronics Research Laboratory at the University of California, Berkeley. Previously, Professor Learned-Miller was the Chief Executive Officer and co-founder of CORITechs, Inc., a company that designed surgical planning software for neurosurgeons.