March 2, 2020

12:00 pm / 1:00 pm

Venue

Clark Hall, Room 316

Seminar 12:00 PM – 1:00 PM
Lunch 1:00 pm – 1:30 PM

                                              

Abstract:  In an idealized imaging model, the optical flow is apiecewise continuous vector field that describes the motion of every pointbetween video frames. The curves of spatial discontinuity of the optical flowfield are called motion boundaries. Detecting these curves accurately helps ina variety of applications including video segmentation or editing and actionrecognition. While methods for computing optical flow from two consecutivevideo frames have rapidly improved over the last few years, their performanceis still lagging along the motion boundaries themselves, making these difficultto detect. The reasons for this difficulty stem mainly from the fact that flowestimators tend to produce smooth outputs, and therefore blur boundaries. Inaddition, motion boundaries have measure zero in the image plane and aredefined differentially, while imagery is discrete in both space and time.

 

I will present a method formotion boundary detection based on neuralnetworks that combines several knowntechniques for pre- and post-processing the input frame pair and pushes thestate of the art on the problem. Our method also includes a novel, simplemodification of the standard encoder-decoder architecture that is used for thistype of problems. This modificationincurs no additional cost, as it merely invertsthe flow of information inthe decoder, and yet yields consistent improvementson a variety of image-to-image estimation problems. This research is joint workwith Hannah Kim.

 

Bio:  Carlo Tomasi received his PhD in Computer Science fromCarnegie Mellon University in 1991. He was assistant professor at Cornell andStanford, and is currently the Iris Einheuser Professorof Computer Science atDuke University. His research spans computer vision from visual motionestimation, image retrieval, and activity recognition to shape reconstruction,stereo vision, texture analysis, and medical imaging. His papers have beencited more than 44,000 times according to Google Scholar, with more than halfof those citations for his top three publications. He won two Helmholtz prizesawarded by the International Conference on Computer Vision for papers that havehad significant long-term impact on computer vision. He is an ACM Fellow, holdseleven patents, and has been principal investigator or co-investigator on morethan 40 research grants.