Weijian Xu

I believe that is due to the data collection process for the original BSDS dataset. You may refer to the paper of BSDS for more details.

Hi @JusticeLin, 1. For error "FileNotFoundError: [Errno 2] No such file or directory: '.\hr\train_pair.lst", please download the data from https://cseweb.ucsd.edu/~weijian/static/datasets/hed/hed-data.tar and make sure the `train_pair.lst` is in the correct path....

Hi @wudemoai, the 5stage-vgg.py36pickle is converted from https://vcl.ucsd.edu/hed/5stage-vgg.caffemodel mentioned in the readme of the official HED repo: https://github.com/s9xie/hed

Yes, it is possible. You may need to modify the test function (see: https://github.com/xwjabc/hed/blob/master/hed.py#L257) and use your own image as input.

The current edge evaluation process in HED is: (1) Generate probabilistic edge maps w/ HED network (i.e., `.mat` files in `mat` folder and png images in `png` folder) -> (2)...

I think you do no need to generate MAT file for the step (3) since its input format is png. You can directly feed the generated png files from Canny...

Well no. Currently the training code only supports batch size 1. So, there is no need to scale the input image.

@peiyunju Yeah I think you can give it a try! Feel free to post some results here if you find such kernel pretty useful.

@caiqi Thx! Will take a look. I am also a newbie to PyTorch and trying to trace the reason of that error.

Got the reason. In PyTorch 0.4, `x.squeeze()[0]` will not return a scalar, but a tensor. It will cause several compatibility problems (e.g. nKbase errors, DAverageMeter errors). Will post a patch...