search

googleinterns / wss

A simple consistency training framework for semi-supervised image semantic segmentation
Apache License 2.0
164 stars 24 forks source link

Question about Tab.5 which compares different pseudo labeling strategies #4

Closed LiheYoung closed 3 years ago

LiheYoung commented 3 years ago

Thanks for your impressive work!

I notice that in Tab.5, different pseudo labeling strategies are compared. Whether the numbers listed here are mIOU on validation set or mIOU of pseudo masks on unlabeled images from the training set?

It seems the latter one, if so, are these unlabeled images from the VOC original training set(1,464 images) or the augmented SBD images?

Yuliang-Zou commented 3 years ago

It is actually the former one. These mIoUs are on the validation set.

LiheYoung commented 3 years ago

But it is somewhat strange that the result(73.79) in Tab.5 of 1/4 labeled setting are higher than that(73.23) in Tab.1 of 1.4k labeled setting?

Is there anything I misunderstood?

Yuliang-Zou commented 3 years ago

That's because they are using different network architectures (and actually different pre-training schedules). In Table 5 we use Xception-65, while in Table 1 we use ResNet. You can find in the original DeepLabv3+ paper that they tune the ImageNet pre-training strategy a bit to get a strong performance for Xception-65.

And I actually used the same hyper-parameters as Xception-65 for ResNets by mistake, which I believe are not optimal. (Due to time constraint, I did not have time to re-train those models using the suggested hyper-parameters in DeepLab repository)

LiheYoung commented 3 years ago

I got it! Thank you very much!

By the way, are all these results produced with single-scale testing, rather than multi-scale testing and flipping?

Yuliang-Zou commented 3 years ago

Yes. All results are tested using a single-scale only, without any test-time augmentations.

LiheYoung commented 3 years ago

Thanks for your prompt reply!