evaluation performances on each category

[zhang-zihui]

Hi,

Thanks for this exciting work and for sharing codes. I want to ask if you can provide the AP results for each category. I know that this paper aims to do a class-agnostic unsupervised object segmentation, but actually we can only use the ground truth for one category to evaluate whether the model perform well on this category, so we can iteratively get evaluation results for all category. Thus I'm wondering would you like to provide such a more finer evaluation results? Thank you again.

regards, zihui

[RozDavid]

Hey @zhang-zihui,

We treated every instance category and a combined foreground class, so we haven't computed per class performance. On the other hand, I do agree this is a pretty interesting metric that we never thought of, I would be also curious to see those results. You could directly modify this script to load per class instance segmentation and compute the matching to all, but keep per class metrics. Pull requests would be highly appreciated.

Cheers, David