About training accuracy about u2pl HOT 7 CLOSED

On which dataset?

from u2pl.

Our checkpoints here for VOC and here for Cityscapes.

from u2pl.

Thanks for providing checkpoint, but I'm more concerned about whether the semi-supervised training can achieve the same accuracy through U2PL.

On which dataset?

Cityscapes.

from u2pl.

On Cityscapes dataset, we apply slide-window evaluation following AEL, which means we should run sh eval.sh after training.
Therefore, the mIoU reported during training is not the final performance, which is quite different from VOC.

Did you run sh eval.sh after training?

from u2pl.

For instance, under 1/2 partition protocol on Cityscapes, the final result is 79.12 with AEL, but the mIoU is around 75 during training.

from u2pl.

By the way, can you reproduce the results on VOC?

from u2pl.

For instance, under 1/2 partition protocol on Cityscapes, the final result is 79.12 with AEL, but the mIoU is around 75 during training.

Thanks. I'll try eval.sh.

from u2pl.