Deep Image Matting v2

Deep Image Matting paper implementation in PyTorch.

Differences

1. "fc6" is dropped.
2. Indices pooling.

"fc6" is clumpy, over 100 millions parameters, makes the model hard to converge. I guess it is the reason why the model (paper) has to be trained stagewisely.

Performance

• The Composition-1k testing dataset.
• Evaluate with whole image.
• SAD normalized by 1000.
• Input image is normalized with mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225].
• Both erode and dialte to generate trimap.

Models	SAD	MSE	Download
paper-stage0	59.6	0.019
paper-stage1	54.6	0.017
paper-stage3	50.4	0.014
my-stage0	66.8	0.024	Link

Dependencies

• Python 3.5.2
• PyTorch 1.1.0

Dataset

Adobe Deep Image Matting Dataset

Follow the instruction to contact author for the dataset.

MSCOCO

Go to MSCOCO to download:

• 2014 Train images

PASCAL VOC

Go to PASCAL VOC to download:

• VOC challenge 2008 training/validation data
• The test data for the VOC2008 challenge

Usage

Data Pre-processing

Extract training images:

$ python pre_process.py

Train

$ python train.py

If you want to visualize during training, run in your terminal:

$ tensorboard --logdir runs

Experimental results

The Composition-1k testing dataset

1. Test:

$ python test.py

It prints out average SAD and MSE errors when finished.

The alphamatting.com dataset

1. Download the evaluation datasets: Go to the Datasets page and download the evaluation datasets. Make sure you pick the low-resolution dataset.

2. Extract evaluation images:

$ python extract.py

3. Evaluate:

$ python eval.py

Click to view whole images:

Image	Trimap1	Trimap2	Trimap3

Demo

Download pre-trained Deep Image Matting Link then run:

$ python demo.py

Image/Trimap	Output/GT	New BG/Compose