COCONut-L are less images than expected

[theo2021]

Hi, I tried creating the COCONut-L dataset and noticed the number of images were different than expected.

I got the COCONut-S -> 118K images.
Then the COCONut-B -> 123K images.
Then for COCONut-L I get from gdrive -> 105835 images.
In total 347526 images. Different from the 358K stated in the paper.

[feivellau]

+1

[xdeng7]

Sorry for the confusion. We remove around 10k noisy labels for the coconut-large object365 subset as they may affect the model performance. Now you should check the huggingface, this will give a collection of object365 subset of panoptic masks, both the panseg mask and infor has been updated. The converted instance masks will be updated soon.

[xdeng7]

Hi, I tried creating the COCONut-L dataset and noticed the number of images were different than expected.

I got the COCONut-S -> 118K images. Then the COCONut-B -> 123K images. Then for COCONut-L I get from gdrive -> 105835 images. In total 347526 images. Different from the 358K stated in the paper.

please noted that, the img_id could be customized into continuous number as we don't use the filenames in object365, you can find the details in the json files.