How to get the categories of the dataset
Thanks for releasing the awesome dataset! I am trying to download the dataset by categories and train my model using only a specified category. I found that there is a "categories" key in annotations, but it is an empty list. Is there a way to organize the downloaded data by categories? Further, is is possible for us to download a subset of the data by specifying categories? What's more, I also found the text descriptions in annotations are lacking. I hope to know if the text descriptions will be released later. And really appreciate your help in advance.
Hi @HJY123 ,
I am trying to download the dataset by categories and train my model using only a specified category. I found that there is a "categories" key in annotations, but it is an empty list. Is there a way to organize the downloaded data by categories?
Some objects do not have categories. See the LVIS annotations as well for categories that are a bit more cleaned up.
Further, is is possible for us to download a subset of the data by specifying categories?
Yes, get all the object IDs corresponding to a particular category, then call load the objects with objaverse.load_objects([uids]). This will only download the subset of uids that you specify.
What's more, I also found the text descriptions in annotations are lacking. I hope to know if the text descriptions will be released later. And really appreciate your help in advance.
Have you looked at the "name" key in the annotations in the annotations? Those are the most succinct text descriptions that we've found.
Thanks for your response. But I am still confused about the LVIS annotations. More specifically, there are only 46k data that got LVIS annotations, but in total there are 760k data. What's more, data in one category(LVIS annotation) is at most 453, which might be not sufficient for model training. And in your paper, I notice that there are 22k categories total while LVIS annotations only contain 1k categories. As for the "name" key, I random sampled annotations and print its "name" contents as follows:
Piu Piu - Molang
очки
Hospital room two
Very much WIP
Basecontelaprueba
Jastew
Sandro
It seems that the "name" contents are hard to understand even for humans, will you improve the text descriptions later? Really appreciate your help!
Yes I have the same question, so I wonder for the remaining non-LVIS annotated data, do we only have those nosiy labels like what HJY123 listed?
Hi @HJY123 and @cnut1648 , thanks for clarifying!
What's more, data in one category(LVIS annotation) is at most 453, which might be not sufficient for model training. And in your paper, I notice that there are 22k categories total while LVIS annotations only contain 1k categories.
Yes, there are only ~1200 LVIS annotations, and at most 500 annotations for each category. (Check the paper details for more on this process.)
LVIS annotations are much more cleaned up than the CLIP category annotations (which is what we got the 22K categorical coverage from). I wouldn't really rely on these 22k categories for much beyond coverage at the moment though.
I'll add in Oscar to this discussion, who may be able to provide them!
It seems that the "name" contents are hard to understand even for humans, will you improve the text descriptions later? Really appreciate your help!
Yes, we are currently working on better text descriptions, and will release them soon!
Thanks for releasing the awesome dataset! So how can I find the CLIP category annotations? I noticed that there are some vehicle models that I need. Could you please tell me how to filter the vehicles from 80k models ? @mattdeitke