Reproducibility concern. about afa HOT 3 CLOSED

The performance I reproduced 63.0 is lower than AA&LR ACM MM’2021 (63.9 VOC2012 valid announced in paper).

from afa.

For experiments with reg loss, the log and weight files can be found at #3 (comment). The full test code is now available here. I did use multiscale and flip inference before CRF. They're common post-processing steps for semantic segmentation, so I didn't expatriate the details in paper.
Since I didn't save the weights without reg loss, I repeated the experiments and got close results again. The new log and weights of the reproduced results are available here (log & weights).
I don't what's wrong with your experiments, but it looks like the reproduction is fine.

from afa.

Thank you for your patient reply. I'll try it soon.

from afa.