7
$\begingroup$

If I look at one of the many sources for the Imagenet classes on the Internet I cannot find a single class related to human beings (and no, harvestman is not someone who harvests, but it's what I knew as a daddy longlegs, a kind of spider :-). How is that possible? I would have at least expected a person class, and even something more specific such as man, woman, toddler, etc. Nothing of the sort. Why? Did Fei-Fei Li and her team make a conscious choice not to have people images in the database? Am I looking at the wrong file? For the sake of the question, we can consider the ImageNet versions from 2014 onwards.

$\endgroup$
4
$\begingroup$

You can also take a look at here for the labels in the imagenet. I guess you are right, there is no label for human in the data-set but there is something to notice. There are labels in imagenet like cowboy or some specific hats and other related things to human like shirt and t-shirt. You can take a look at here and also here. In the latter link Yosinski et al, has tried to show that the popular AlexNet has learned to recognize human faces although there is no label as human face in the imagenet data-set. In their paper, they have investigated that Convolutional neural networks may try to learn things that are distributed among layers or maybe not and they may not have special label in the training data. As an example, the face of cats and humans can be referred to. Moreover, as you can see here maybe the aim was attribute learning in large-scale datasets, as quoted in the last line of the page, as the reference.

$\endgroup$
  • $\begingroup$ I like the answer, but I don't understand your last sentence. What do you mean by attribute learning, how it differs (if it differs) from image classification and how is this related to my question (are there person classes in ImageNet)? $\endgroup$ – DeltaIV Feb 11 '18 at 10:21
  • $\begingroup$ @DeltaIV I meant in the last link that I've provided, there is a reference which in that work, they discuss this issue. What I said was like learning faces, which are not the labels but are needed to understand t-shirts maybe. $\endgroup$ – Media Feb 11 '18 at 10:47
  • $\begingroup$ Ok, the NNs learn features which look like faces because they help recognizing (or discerning between) labels. Yes, I kind of expected that. Thanks $\endgroup$ – DeltaIV Feb 11 '18 at 10:55
  • $\begingroup$ @DeltaIV I guess this is maybe what we call learning $\endgroup$ – Media Feb 11 '18 at 11:01
  • 2
    $\begingroup$ I think learning for neural networks has very little to do with the human learning process. See these hallucinations. Then again, the same idea that these optimized images should show what Neural Networks have learnt, is deeply flawed and based on a misunderstanding of what an high-dimensional probability distribution is. The subject is very delicate: my question was much more elementary. $\endgroup$ – DeltaIV Feb 11 '18 at 11:16
4
$\begingroup$

I found the class 7846 (name="n00007846") is for person. To access to class description, read http://image-net.org/download-API. Even better, the following text files contain everything you ever need to understand the classes in ImageNet dataset (class = WordNet ID):

n00007846 maps to person, individual, someone, somebody, mortal, soul. The corresponding gloss is: a human being; "there was too much for one person to do".

$\endgroup$
  • $\begingroup$ Could you link to a source? That would be helpful for other users. $\endgroup$ – Elias Strehle Mar 9 '18 at 12:33
  • $\begingroup$ Imagenet models seem to be underperforming on humans of various backgrounds. In my data there are people in running outfits and they mostly identified as rugby balls and volleyballs. $\endgroup$ – levesque Feb 18 at 20:47

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.