Skip to content

Comment on Large Scale Visual Recognition Challenge 2011 - Resultsparent

Comments

So you're asserting that the 10% improvement by Supervision is because they used the raw RGB pixels. Is that right?

If so, then I'm guessing that the other teams only used compressed representations like Fisher vectors with linear classifiers, because they needed to scale. Instead, Supervision achieved scale with raw power, doing the training computations on GPUs for a week (probably coded in openCL).

"So you're asserting that the 10% improvement by Supervision is because they used the raw RGB pixels. Is that right?"

No, what I meant is that SuperVision did very well because their feature space is richer than the other teams, but IMHO that is resulting for the deep convolutional process which it used to generate rich features. This is a good explanation of the subject:

http://ai.stanford.edu/~ang/papers/icml12-HighLevelFeaturesU...

I deleted all the comments I could as they were complaining about the title. Since it was changed to a proper one most of my comments are not relevant any more.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.