tulips

ICCV - Day 4

Now that I’m free from my poster duties (but by no means, very tired from talking and standing), I got to take it easy, enjoy some good talks, pick up new ideas, and go check out some other posters and demos.

This year’s David Marr Prize for the best paper went to Devi Parikh and Kristen Grauman on Relative Attributes, an interesting concept of formulating training data for object/scene classification based on how object/scene categories relate according to different attributes, as opposed to strictly binarized prediction of attributes. It all sounds pretty much like a re-visitation of ideas from fuzzy logic but in a much more elegantly structured model, if you ask me.

Some other interesting oral sessions I clearly remembered (!):

• Human Action Recognition by Learning Bases of Action Attributes and Parts, by Bangpeng Yao et al.
• Describing People: A Poselet-Based Approach to Attribute Classification, by Lubomir Bourdev et al.
• Discovering Favorite Views of Popular Places with Iconoid Shift, by Tobias Weyand (who happened to have some really cool personal projects) and Bastian Leibe

Saw a really cool demo on real time Dense Tracking and Mapping (DTAM) by the people from Imperial College London (first picture above). Pretty impressive stuff. Very precise and quick even with a heavily cluttered mountain of objects. They also have a video demo in YouTube which you can watch over here.

And…they installed a pair of lights for every poster board on the 2nd day, which sort of brightens up the poster and the whole atmosphere. Not fair! But from what I gather, many of the presenters actually did not really like the lights as it obstructed part of the poster from the frontal view, and that the extra heat generated can be unbearable after awhile…