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:

PASCAL VOC

Go to PASCAL VOC to download:

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
  1. Evaluate:
$ python eval.py

Click to view whole images:

Image Trimap1 Trimap2 Trimap3
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image
image image image image

Demo

Download pre-trained Deep Image Matting Link then run:

$ python demo.py
Image/Trimap Output/GT New BG/Compose
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image
image image image

GitHub