I have seen an error in the file fcn8s_pascal.yml, in the selection of the dataset.
For the val split it select "val" instead of "train_aug_val" this leads to 400 images that are both in training and validation set.
After correction, the accuracy of the FCN8s network fall from 65% to 57%.
I think it is important to correct it, and do you know how to get an implementation that achieve the same as the paper?
Thank you in advance
hjhjb
commented
1 year ago
I have seen an error in the file fcn8s_pascal.yml, in the selection of the dataset. For the val split it select "val" instead of "train_aug_val" this leads to 400 images that are both in training and validation set. After correction, the accuracy of the FCN8s network fall from 65% to 57%. I think it is important to correct it, and do you know how to get an implementation that achieve the same as the paper? Thank you in advance