Can not reimplement the mIoU in the paper.

[Zxl19990529]

Hi, thanks for your work. I'm currently following your work. However, I can not reimplement the result in the paper. My reimplementation is just 70.95. The followings are my reimplementation:

1. I use the this repo to generate the scribble masks.
2. I found the resnet101 model in current repo is different from the original one. In original resnet101, the conv1 is a single convolutional layer with kernel size 7*7, stride 2 and output channel 64. In current repo, the conv1 is a "Conv block" with conv(3,64,3,2,1)-bn-relu-conv(64,64,3,1,1)-bn-relu-conv(64,128,3,1,1). As I found no pretrained model released, I replace the current one to the pytorch offical res101 model, and load the pretrained model from pytorch model zone.

[Luffy03]

Thanks for your kind comments.
The pre-trained models and weak labels are released now. Specifically, we use the weak labels provided by TEL.
Feel free to contact me if you have any questions.