tulips

ICCV - Day 1

Pre-conference events started with an attractive bunch of tutorials on Day 1.

• I had in mind to attend the very general, multi-topic, summary-of-field “tutorial” titled Looking at People: The past, the present and the future, but decided at the last minute to sneaked over to the tutorial on inference learning with discrete graphic models. A good brief overview by Ramin Zabih, which quickly turned out to be quite thoroughly specific towards energy models, Markov Random Fields (MRF) etc. Wasn’t as well-structured as I had expected.

• Popped back to the Looking at People tutorial after lunch. Was pleasantly surprised by a couple of interesting (and I should also use the word brilliant here) talks on Activity Recognition, by Aaron Bobick and Xiaogang Wang. Bobick with a historical take on how human action analysis came about, Wang on the state-of-art usage of topic models.

Take-home eye-openers

• Topic models: Latent Dirichlet Allocation is a very powerful generative model that can be used to characterize spatio-temporal dependencies from among the many variables (observed variables and latent variables) in a concise manner. Another topic model, Probabilistic latent semantic analysis is similar, but doesn’t use Dirichlet priors.

• If I had only one thing to comment, it would be how utterly research-oriented, passionate and geeky people are over here. Main event has not even begun, and the first day tutorials were already packed to the brim and everyone’s either attending talks, banging away at the laptop, or talking about research. This is HOW a conference should be.