clovaai / wsolevaluation

Evaluating Weakly Supervised Object Localization Methods Right (CVPR 2020)
MIT License
332 stars 55 forks source link

About the class of objects #61

Closed Breezewrf closed 9 months ago

Breezewrf commented 9 months ago

Hi,

I wonder why the class label annotated like this: 001.Black_footed_Albatross/Black_Footed_Albatross_0009_34.jpg,0

And it is weird, all images in your dataset(CUB, I just downloaded the small one dataset) only include single object?

I am new to wsod, is it usual to train the model in this style like only detect single object rather than multiple objects?

Thank you for your answer!

SanghyukChun commented 9 months ago

https://github.com/clovaai/wsolevaluation/blob/96f4b88679af41f005824661f47e6fe3628af2aa/data_loaders.py#L50-L83

The metadata is csv where each column denotes image id and class name. Hence, 001.Black_footed_Albatross/Black_Footed_Albatross_0009_34.jpg is the image id, and 0 is the class label.

WSOL only contains a single object, and no object box is provided. We only train a normal classifier and extract heatmaps using some techniques (e.g., CAM). If you would like to work on WSOD, you'd better to search another repository for WSOD (detection), not WSOL (localization)

Breezewrf commented 9 months ago

I got it, thanks a lot!