Why isn't there a person class at ILSVRC'12 classification task? Recently I played a bit around with the Caffe Classification Demo and recognized bad classification results with images of people. I guessed the model is trained with the ILSVRC'12 classification dataset. When looking at it in more detail I realized there isn't a human/person class. The most related classes are just "scuba diver", "groom" and "ballplayer, baseball player". 
So why there are added many classes for fine-grained classification, but there isn't a simple person class? I couldn't find a direct explanation at the ImageNet paper. Isn't it impractical for many applications?
 A: I think there are two likely reasons (but I'd be delighted if you email the researchers!):
Firstly, categories in underlying ImageNet dataset are derived from WordNet, an older NLP dataset of basically synonyms (synsets). WordNet consists of about 150,000 words arranged in an essentially hierarchical order (you can browse WordNet here).  Part of the goal of the creation of the ImageNet was not just to build a very large image dataset but also to have the images arranged in a dense semantic hierarchy with around 500-1000 images per leaf node in the hierarchy. ImageNet was originally based on the vehicle and mammal subtrees.  If you browse through WordNet, I think it becomes clear that not all the trees are related to objects with distinct visual properties. If you look at the synset "firemen", you'll see it's part of the follow pretty abstract hierarchy:

fireman -> defender -> preserver

There is an hierarchy for person in ImageNet, but I think it suffers from the same problem of being quite abstract, whereas the hierarchy for dogs is a little more concrete.  You can find more information in the paper by Deng et al. 2009 "ImageNet: A Large-Scale Hierarchical Image Database" on the ImageNet dataset (you are citing the paper about ILSVRC which is different). That's my guess though about why persons was not included in ILSVRC.
Secondly, working with images and video of people is almost it's own subdiscipline in computer vision. There are specialised models and subtasks in like face detection, activity recognition, pedestrian detection, and eye tracking. If the researchers were interested more generally in object recognition problem, they might not want deal with all the complications of dealing with people.
