Open davidwagner opened 3 years ago
Yeah, so extra/*/*
is an unfiltered dataset that will have a lot of noise but it's designed to have as images as possible so that we can train useful classifiers. (For example, if you use an imagenet classifier as a feature extractor, then a single linear projection layer can get 99% test accuracy by training on extra.)
I believe (in theory) that train/*/*
and test/*/*
should satisfy the constraints. We're going to (very shortly -- early next week?) release a new version of the dataset that will rename current test to val, and then introduce a new test dataset with in-the-wild images that will be a subset of the held-out test set to be used in the full contest.
The contest provides three sets of images: train, test, and extras. Which of these have already been verified by taskers to meet the criteria to be a valid unambiguous image? It might be be useful to document this somewhere.
Looking through 40 images in extras chosen at random, it looks like these have not been filtered. Examples: