Clay Gong

Yes, you can change the first layer and use the placeholder to input images.

Yes. You can refer to our example images. Or just download the edge images and change the name.

labels/*.png should have only one channel. Have a try?

We didn't have this problem. Maybe you can check the format again? We use MATLAB R2016a on Ubuntu 16.04

@makalo I don't know which two train_ids you have downloaded. Can you provide the detailed path?

The input size doesn't matter. I think the method fails for this kind of images when the person doesn't face directly at the camera. Maybe you can try the more...

Actually, the inference time depends on your input size. It will be slower for the multi-scale inference.

There are many reasons for the poor re-trained results, such as pre-trained model, training epochs. We use sum operation for edge loss to follow the baseline, _Holistically-Nested Edge Detection_

For difference, PGN is totally a framework for parsing images while ATEN aims at utilizing temporal information for video parsing. For frame-level accuracy, PGN is better while for sequence input,...

http://www.stat.ucla.edu/~xianjie.chen/pascal_part_dataset/pascal_part.html#statistics