Full scripts for training

[lrjj]

Hello，could u please provide a few samples of your training data and the full scripts for training? Thanks

[ZHKKKe]

Hi, thanks for your attention. Releasing of our full scripts for training is scheduled after this paper is accepted.

[kapulkin]

I read the training code. Method supervised_training_iter takes image, trimap and gt_matte.

To generate trimap I wrote next code:

def makeEdgeMask(mask, width):
    kernel = np.ones((width,width), np.uint8)

    erosion = cv2.erode(mask, kernel, iterations = 1)
    dilation = cv2.dilate(mask, kernel, iterations = 1)

    return dilation - erosion

def makeTrimap(mask, width = 5):
   edgeMask = makeEdgeMask(mask, width)
   trimap = mask.astype(np.float)
   trimap[edgeMask == 1] = 0.5
   return trimap

@lrjj, may be that will help you.

In my case code training technically works - semantic, detail and matte losses become very low:

But generated masks are incorrect.

As training data I took 640x480 CocoDataset images.

[vkl-git]

I read the training code. Method supervised_training_iter takes image, trimap and gt_matte.

To generate trimap I wrote next code:

def makeEdgeMask(mask, width):
    kernel = np.ones((width,width), np.uint8)

    erosion = cv2.erode(mask, kernel, iterations = 1)
    dilation = cv2.dilate(mask, kernel, iterations = 1)

    return dilation - erosion

def makeTrimap(mask, width = 5):
   edgeMask = makeEdgeMask(mask, width)
   trimap = mask.astype(np.float)
   trimap[edgeMask == 1] = 0.5
   return trimap

@lrjj, may be that will help you.

In my case code training technically works - semantic, detail and matte losses become very low:

But generated masks are incorrect.

As training data I took 640x480 CocoDataset images.

Hello, if it is convenient, could you please send me a copy of your training script?Thank you very much.

[Fight-hawk]

哈喽 我实现了一个完整的训练代码 用的是bg的数据 效果不好 欢迎一起探讨一下