Hi,
Thank you so much for such a wonderful work.

Can this model be used for getting real-time alpha mattes for images other than portraits as well? Like, Would the outputs be good if we trained it on Deep Image matting dataset?

Hi,
I see in the paper that higher depth multiplier (0.5, 0.75, 1.0) leads to lower gradient error. My question is what is it and how much i can increase it?

@shurain 안녕하세요. 좋은 소스를 공개해주셔서 감사합니다.
이모델을 트레이닝 하려고 하는데, 데이터셋의 구성을 어떻게 해야 하는건가요?

제안하신 디렉토리 구조는 이해가 됩니다.

dataset_directory
|___ train
| |__ mask
| |__ image
|
|___ test
|__ mask
|__ image

이때 원본 이미지와 마스크 이미지의 형태를 어떻게 구성하면 될까요?

1. 이미지의 포맷은 어떻게 지정합니까? png, jpg 상관없나요?
2. 마스크의 영역의 구성을 확인하고자 합니다. 마스크는 grayscale 로 bg 0, shape 1 로 칠했는데, 맞을까요?

그리고 트레이닝이 끝나는 조건이 있나요? 너무 오래걸려서요

GPU Tesla K80 입니다.

image

mask

그리고, input / output node 명좀 먼지 알려주세요. freeze_graph 할려고 합니다. 혹시 covert_to_pb 설정이 가용한건지 확인도 부탁드립니다.

감사합니다.

We went through your model. The outputs were surprising. Could you share with us your pre-trained model? Is the model giving these results in Mobile CPU?? You have handled quantized models, but how did you manage to preserve the accuracy to good extent.

not test out mask ?

When I train, the log will output in the terminal, but I can't find a log_file in the project path? Is there a log_file for reviewing?

This "--convert_to_pb" in evaluator.py but no relevant codes to do it?

Is there an option which can continue-train if I have to break and I want to train continuely from last training coondition?

authors, thanks for opensourcing your code.
I would like to know why did you choose tf.contrib.slim instead of tf.layers or just tf.keras?
I'm trying to reimplement your using tf.keras which could be more simpler for more expremintation with tf.eager_execution to debug easily? what are implecations I should take to use tf.keras?
Thanks!!

We tried out your model. The output quality is good. We referred your paper, wherein it mentions that you got good result in tflite level. We tried out your pb file output, and the output is coming good. When we tried out the tflite conversion, the tflite is created and output is not coming properly. There are huge changes in the weights between the frozen graph convolution operation and tflite convolution operation when the tflite conversion happens. We are using the fused batch norm approach. Need ur help

Can you tell me where model is if i want to run demo?

I found it train on CPU

Why change

Before: [SpaceToBatchND] -> [DepthwiseConv2dNative] -> [BatchToSpaceND] -> [normalize] -> [activate]

to

After : [SpaceToBatchND] -> [DepthwiseConv2dNative] -> [normalize] -> [activate] -> [BatchToSpaceND]

Hello
How are you?
Thanks for contributing this project.
Could u provide the inference script for an image?
Thanks