remove small objects about pfenet

[tianzhuotao]

Hi,

Thank you for being interested in our work.

The preprocessed ground truth images of CANet are actually processed by the scripts of OSLSM and therefore the small objects are also removed in the project of CANet (we have confirmed with the authors of CANet before we submit our paper).

Also, I think it is okay to follow OSLSM to screen out small objects (we only remove them for support samples) because OSLSM is the first paper that proposes the setting of few-shot segmentation.

Frankly speaking, it is hard to keep all papers to have the same configuration. If you want to have a comparison, you can try to remove the small objects for support samples using the repo of PANet or keep them in our repo to see the performance gap it causes to the model.

Thank you and hope my explanation is helpful.

[LiheYoung]

I think it is more practical to only remove them for support samples. Thanks for your explanation!

[LiheYoung]

Sorry to bother you again. But do you mean OSLSM and CANet remove small objects in both support and query set?

[tianzhuotao]

@LiheYoung I think the small objects are removed in CANet. However, from my perspective, we do not want to alter the data sample for validation (query samples), even if some objects are small and hard for models to identify, therefore we only remove them from support samples.

[LiheYoung]

@tianzhuotao Thanks for your explanation!

[LiheYoung]

I take a closer look at the dataloading process.

It seems that if an image doesn't contain any objects from current class space (such as base/novel class) larger than 2x32x32, the image will be removed from the data list. This image will never be selected during current phase. And the small query objects in these removed images will never be sampled. Therefore, actually, it still removes a portion of small query objects, which may bring slight better performance.

Is there anything I misunderstand? Thanks!

[tianzhuotao]

@LiheYoung

Thank you for mentioning it!

I overlooked this point when I reviewed my code. Yes, a portion of small objects in the query set has been excluded in my implementation, which causes very little improvement on average in my case. But it still has an overall fair comparison with other methods I think.

However, this repo can achieve higher performance than the results reported in my original paper (see this issue). Of course, for a fair comparison, you can accordingly modify my code to be the same as that in your setting.

Thank you again! Please drop me a message if you have further questions.

[LiheYoung]

@tianzhuotao Thanks for your reply!

[cty21995]

Why my test result of moiu is 0