zero-shot

[yujialele]

How to split rare first datasets and non-rare datasets on HICO-DET. And please provide Unseen object list for me

[zhihou7]

Hi @yujialele ,
the unseen HOI categories of non-rare and rare first splits are provided in https://github.com/zhihou7/HOI-CL/blob/a23c77a410c7e3fcffbf000e87ff5fa84cc6c963/lib/ult/ult.py#L2028.

I first sorted the categories list according to the number of instances. Then I select the HOI categories from the head and tail as non-rare and rare first splits respectively. Particularly, I make sure that the factors (verb and object) of unseen HOI factories exist in the remaining categories( seen HOIs) during selection. that is, if I find there will no instances for one verb or object in the remaining HOI categories, the current HOI category will be kept, and I check the next categories for selecting as unseen categories.

For the unseen object list, it is {'toothbrush', 'frisbee', 'elephant', 'zebra', 'baseball_bat', 'skis', 'tennis_racket', 'refrigerator', 'skateboard', 'dog', 'pizza', 'clock'}. You can check it in unseen hoi list.

The unseen object splits follows https://arxiv.org/abs/1904.03181. Please notice this split will also result in unseen verbs according to my comment in this line. that is ['break', 'slide', 'move', 'flip', 'brush_with'].

Feel free to ask if you have further questions.

[yujialele]

Thanks! I want to know how the unseen hoi category is defined in your work. I read this paper named Detecting Human-Object Interactions via Functional Generalization(https://arxiv.org/abs/1904.03181), but I find that they randomly select 12 objects. In your paper FCL, whether the unseen objects are randomly selected or not.

[zhihou7]

No，unseen objects are from their selection. It is the same as cosnet. You can treat that the unseen categories are fixed in my paper. However, As I mentioned before, current unseen object split is not much reasonable. Get Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ From: yujialele ***@***.***> Sent: Wednesday, October 26, 2022 6:49:48 PM To: zhihou7/VCL ***@***.***> Cc: Zhi Hou ***@***.***>; Comment ***@***.***> Subject: Re: [zhihou7/VCL] zero-shot (Issue #10) Thanks! I want to know how the unseen hoi category is defined in your work. I read this paper named Detecting Human-Object Interactions via Functional Generalization(https://arxiv.org/abs/1904.03181), but I find that they randomly select 12 objects. In your paper FCL, whether the unseen objects are randomly selected or not. — Reply to this email directly, view it on GitHub<#10 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AQGPLYHSKDRNFVQIFXPBVJTWFDPBZANCNFSM6AAAAAARK2K2AI>. You are receiving this because you commented.Message ID: ***@***.***>