Dear all,

Dr. Matthew Grimes from Cambridge University will be visiting the Robot Vision group on 28th November, and giving a talk at 1 pm. The abstract of his talk and his bio are in the following.

If you are interested in meeting him, please drop me a line. I’ll arrange this.

————————————————–
Date: 28 November 2014, Friday
Time: 13:00 – 14:00
Room: Huxley – Room 218
————————————————–

Title: Towards live classification with convolutional networks

Abstract:

Convolutional neural networks have excelled in standardized image classification benchmarks such as ILSVRC, where millions of crowd-labeled images are used to train 1000-class classifiers.

Performing recognition and localization on live video poses its own challenges. Live video can present oddly-framed views not represented in datasets captured by photographers. Recognition speed, a concern ignored in many benchmarks, is at a premium. The cost of building a custom training dataset must be minimized. Finally, industrial settings can put a higher premium on the ability to make fine-grained distinctions (e.g. “spanner” vs “adjustable spanner”).

I will present work in progress on view-invariant classification and localization of objects in video, motivated by previous work on using convnets for live terrain classification for ground robotics. We demonstrate robustness to occlusion, the ability to distinguish similar objects, and invariance to pose, lighting, and background.

Bio:

Matthew Koichi Grimes is a post-doctoral research associate in the Computer Vision & Robotics Group, Machine Intelligence Lab, at the University of Cambridge. Prior to this, he was the localization and mapping lead at the Autonomous Driving group at the Bosch Research and Technology Center, working on laser and vision-based mapping and localization for fully automated highway driving with lane changes. He did his Ph.D. at New York University under the supervision of Yann LeCun. His dissertation was on real-time city-scale SLAM, and fast hybrid methods for outdoor visual odometry.