Architecture

Generator

A decoder encoder based architecture is used.

There are 2 options for the generator encoder.

a. Resnet50 minus the last 2 layers b. Resnet50 + ASPP module

The Decoder network of the Generator network has seven upsampling convolutional blocks. Each upsampling convolutional block has an upsampling layer, followed by a convolutional layer, a batch normalization layer and a ReLU activation function.

Discriminator

The discriminator used here is the PatchGAN discriminator. The implementation here is inspired from the implementation of CycleGAN from
https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

Again, 2 different types of discriminator are used a. N Layer Patch gan discriminator, where the size of the patch is NxN, it is taken as 3x3 here b. Pixel patch Patch gan discriminator, the discriminator classsifies every pixel.

How to use

Use the dataroot argument to enter the directory where you have stored the data. Structure the data in the following way.

train
-alpha -bg -fg

test
-fg -trimap

The background I have used here is the MSCOCO dataset.

To train the model using Resnet50 without ASPP module

!python train.py --dataroot ./ --model simple --dataset_mode generated_simple --which_model_netG resnet50 --name resnet50

To test the model using Resnet without ASPP module

!python test.py --dataroot ./ --dataset_mode single --which_model_netG resnet50 --ntest 8 --model test --name resnet50

To train the model using Resnet50 using ASPP module

!python train.py --dataroot ./ --model simple --dataset_mode generated_simple --which_model_netG resnet50ASPP --name resnet50ASPP

To test the model using Resnet50 using ASPP module

!python test.py --dataroot ./ --dataset_mode single --which_model_netG resnet50ASPP --ntest 8 --model test

Results

Input:

Trimap:

AlphaGAN matting :

Generator:Resnet50,Discriminator:N Layer Patch GAN

Generator:Resnet50,Discriminator:Pixel Patch GAN

Generator:Resnet50 + ASPP module,Discriminator:N Layer Patch GAN

Comparing with the Original implementation

(Average Rank on alphamatting.com has been shown)

Error type	Original implementation	Resnet50 +N Layer	Resnet50 + Pixel	Resnet50 + ASPP module
Sum of absolute differences	11.7	42.8	43.8	53
Mean square error	15	45.8	45.6	54.2
Gradient error	14	52.9	52.7	55
Connectivity error	29.6	23.3	22.6	32.8

Training dataset used

I used the following link for training dataset. Link to created dataset