search

dongwu92 / AutoPortraitMatting

Tensorflow implementation of Automatic Portrait Matting on paper "Automatic Portrait Segmentation for Image Stylization"
Apache License 2.0
478 stars 151 forks source link

Incorrect transform in preprocessing.py #31

Closed MattKleinsmith closed 6 years ago

MattKleinsmith commented 6 years ago

With transform in preprocessing.py

bad2

With correct transform

good

Fix

https://github.com/PetroWu/AutoPortraitMatting/pull/32

Discussion

The bottom center image of each example is the mean mask aligned with the given portrait. They're intended to be rough estimates of what the final mask should look like. The misalignment in the first example is due to using an incorrect transform.

MLjian commented 6 years ago

Hi，I am doing my homework, I have some questions about it. Hope to communicate.

MattKleinsmith commented 6 years ago

@MLjian Sure, what are your questions?

MLjian commented 6 years ago

Did you run the example? How is the result?

MattKleinsmith commented 6 years ago

@MLjian

I'm stuck at 90% IoU on a validation set, while the authors got 96% on the test set. This is with my custom code, though.

I haven't ran this repo's code. I only looked at the code to compare my implementation of the paper with PetruWu's implementation. I noticed an error, and so I posted here.

MLjian commented 6 years ago

You are so passionate. Thank you very much.I saw the paper Automatic Portrait Segmentation for Image Stylization carefully. And what's the architecture in deep learning for implement? Tensorflow or other.

MattKleinsmith commented 6 years ago

@MLjian

The paper uses a custom Caffe build from 2014. I'm using PyTorch.

MLjian commented 6 years ago

Hi,guy! The datasets folder are so complex for me. Now I want to just use the FCN without additinal channels to complete the matting. So I want to know whether the Images_mask folder is the annotation that the paper author labeled.