search

saeedizadi / binseg_pytoch

Implementation of several state of the art methods for Binary Segmentation in PyTorch
Apache License 2.0
18 stars 9 forks source link

DUC_HDC performance #1

Open InstantWindy opened 6 years ago

InstantWindy commented 6 years ago

Hello! Do you use the 'DUC_HDC‘’ model train the Voc2012 dataset? I use it ,but I found the result was not good, it appears the griding situation,But it should not be the case, because the model is the solution to the grid phenomenon.I don't why ,could you help me? Thanks!

saeedizadi commented 6 years ago

Hi,

Yes, I'm getting the grid-like pattern as well. The implementation is correct, and I'm skeptical to the theory behind the DUCHDC method.

InstantWindy commented 6 years ago

Thanks! I have another question. If it is a binary classification, using BCELoss to calculate the loss, how do you calculate the accuracy and mean IoU? Is BCELoss good for image segmentation? I'm sorry to trouble you

saeedizadi commented 6 years ago

BCE stands for binary cross entropy. So, it's suitable for binary segmentation task. The accuracy and IoU can be simply calculated for binary segmentation tasks. Just treat each pixel as a sample. Yes, BCE is good for binary segmentation.