Peekaboo

After looking trough ImageNet for a little while now, I found some things that I did not really expect. So here are some properties of ImageNet that I found interesting (even though some of them might be obvious). But first, a quick recap on what ImageNet is: It's a hand annotated dataset, consisting of 10 million images with 10 thousand object classes. The images were collected using search engines and flickr. Classes correspond to "synsets" in WordNet. A synset is a collection of semantically equivalent nouns. For example, there is a synset called 'n04037443' (this is the IMID, the image net id), which corresponds to the nouns 'racer, race car, racing car' and is described as 'a fast car that competes in races'. The synsets in WordNet have an additional hierarchical structure, given by a directed graph. Going down the graph goes from more general concepts to more specific concepts. For example 'mammal' is above 'canine' whi...

I guess it is a bit late now for starting to work on image net, since the ILSVRC competition just ended. Still, I think this is a very interesting - if not the most interesting - dataset available today in computer vision. So I thought I'd finally take a look. As usual I wrapped up some Python scripts to do the work for me. You can find my little script on github . At the moment, I provides a class to parse the meta data and xml annotations. It can produce bounding box images and you can search for synsets by keywords and browse the tree somewhat. You have to download the dataset and annotations yourself, though. I am still working on the code and additional functionality is likely to be added soon.

Label Embedding Trees for Large Multi-Class Tasks by Samy Bengio, Jason Weston, David Gran is a paper about coping with classification in the presents huge amounts of data and many many classes. Of course the image net challenge comes to mind. But what they actually did is doing classification on all of image net , which is about 16.000 classes and over a million instances. This work focuses on very fast recall but is very expensive to train - apparently the model was trained in a Google cluster. This paper explores two ideas: hierarchical classification and low dimensional embeddings. For the hierarchical classification part, a tree of label sets is build where the leaves are single classes. Starting from all classes at the root, the set of classes is split into subsets similar classes which are represented by new nodes. This is done by training a one-vs-rest classifier and inspecting the confusion matrix. Classes that are easy to confuse are put into the same node. At test-...