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

A doubt #2

Closed yuanyuwan closed 7 years ago

yuanyuwan commented 7 years ago

Good job! I'm also interested in these two paper ("Automatic Portrait Segmentation for Image Stylization" and "Deep Automatic Portrait Matting"). But I'm not sure whether the method of matting proposed in these two paper are same. Could you tell me your opinion?

dongwu92 commented 7 years ago

The second paper（denoted as eccv16）uses the first paper（denoted as eg16）'s network as inputs of trimap. But in eccv16, the author uses three d×d×d×d tensors to predict, while in eg16 he uses softmax to directly predict. This should be the main difference. I implemented eccv16 but the three tensors are too large that i can train it on my device.

yuanyuwan commented 7 years ago

I agree with you. The three tensors correspond to foreground, background and unknown pixels. Thanks.

dongwu92 commented 7 years ago

The diag matrices of F/B and the Laplace matrix key of eccv16, but they are too large and it has limited contribution to the community.