The second day of CVPR13 continued with papers, talks and demos by the world's foremost researchers in computer vision. Read on to learn about some of today's happenings, and notably for us, the announcement of the recipient of the CVPR13 Best Paper Award!
-
We continued to run Gesture Pressure! at the Google booth, a game that demonstrates the recent enhancements to the Android keyboard: contestants are given a set of phrases to enter via either gesture typing or index finger tap input; the contestant with the most entries in 1 minute win the Nexus 7 at the conclusion of the conference. Also at the booth was Google Software Engineer +Jiajun Zhu, answering questions on the Google Self-Driving Car, as well as eagerly speaking with those interested in working on robot perception and computer vision problems. 

Later in the day, the Google booth showcased Google Play Movies (http://goo.gl/zlw1n), highlighting the recent addition of Info Cards to the Google Play Movies & TV app. With Info Cards, activated by pressing the pause button on your tablet, one can easily learn more about the actors, related films, and even what song is playing in a scene of a movie. 

During the oral sessions, Queen Mary University of London EECS PhD student Ravi Garg presented Dense Variational Reconstruction of Non-Rigid Surfaces from Monocular Video, highlighting research focused on accurately reconstructing three dimensional information for pixels in videos of deformable surfaces (e.g. cloth, a human face, or a beating heart). While obtaining 3D models of a scene observed by a camera is a key problem in computer vision, currently successful methods of Structure from Motion (SfM) are limited to handling scenes containing only rigid objects, additionally requiring fully calibrated images and prior models or shape templates.

Working with co-authors Anastasios Roussos (http://goo.gl/Mxx9d), and Lourdes Agapito (http://goo.gl/4CQOm), Garg described the method by which they model deformable surfaces from a single video sequence, with no markers or pretrained shape models, formulating a NonRigid Structure from Motion (NRSfM) algorithm as a global variational energy minimization problem. With NRSfM, the authors show that this approach successfully generates dense 3D reconstructions by utilizing terms that minimize image projection error as well as spatial regularization terms that provide smooth 3D shapes, presenting the first variational approach for dense 3D reconstruction of non-rigid scenes from a monocular video sequence without prior scene knowledge or shape templates. To get the further details, read the full paper at http://goo.gl/ZQJwl.

In the “Detection” oral session, Google Software Engineer Sudheendra Vijayanarasimhan presented the recipient of the CVPR13 Best Paper Award, Fast, Accurate Detection of 100,000 Object Classes on a Single Machine, co-authored by Googlers Thomas Dean, +Mark Ruzon, Mark Segal, Jonathon Shlens, and Jay Yagnik. In the talk, Vijayanarasimhan addressed the computational time constraint limiting current object detection systems, which convolve a target image with a bank of filters that code for different aspects of an object’s appearance, with the goal of localizing all the objects in a an image and recognize all the categories they belong to. By utilizing locality-sensitive hashing (http://goo.gl/ymbkP), which essentially reduces the dimensionality of the filters by replacing them with a fixed number of hash-table “probes”, the authors demonstrated an speed increase of over four orders of magnitude over current object detection algorithms. 

Not restricted to any particular method or dataset, the authors’ approach can be applied to a variety of signal processing tasks related to image processing, and provides a method that reduces the dependence of the detection time on the number of classes. By introducing a scalable approach to object detection with a locality-sensitive hashing scheme, the authors show how this method is able to handle millions of convolutions and simultaneously detect up to 100,000 object classes on a single machine in under 20 seconds. Read the full paper at http://goo.gl/SGj83

The second day ended with the announcement of the CVPR13 Best Paper awards. Google is proud to sponsor the Best Student Paper Award, given to Discriminative Non-blind Deblurring, authored by Uwe Schmidt, Carsten Rother, and Sebastian Nowozin from the Department of Computer Science at TU Darmstadt, and Jeremy Jancsary and Stefan Roth from Microsoft Research Cambridge. Read the full paper at http://goo.gl/tjrY8. Recipient of the Best Paper: Runner-Up Award was Lost! Leveraging the Crowd for Probabilistic Visual Self-Localization, authored by Marcus A. Brubaker and Raquel Urtasun of TTI Chicago, and Andreas Geiger of KIT & MPI Tubingen. Read their paper at http://goo.gl/5Injv.  And of course, congratulations to the Googlers who worked hard on the CVPR13 Best Paper, described above.

Tune in again tomorrow, for coverage on the final day of CVPR13!
Photo
Photo
2013-06-26
2 Photos - View album
Shared publiclyView activity