search

kvmanohar22 / DIM

Deep Image Matting
3 stars 0 forks source link

readme

Deep Image Matting

This is Chainer implementation of the paper Deep Image Matting

Example Results

Evolution of matte as training progresses

For other experimental results check them out at the bottom of this page.

Installation

It's recommended to use anaconda environment.

Requirements

  1. Install chainer and cupy from here
  2. Install the other requirements using pip

Usage

Directory structure

/path/to/DIM
         |-> ckpt/
         |     |---> demo.npy           # Checkpoint files are loaded from here
         |-> nnet/
         |     |---> model.py           # Builds graph of the neural network
         |-> scripts/
         |     |---> train.sh           # Training script
         |     |---> demo.sh            # Demo script
         |     |---> test_functions.sh  # Testing utils script
         |-> tests/
         |     |---> test_utils.py      # Tests for utils
         |-> utils/
         |     |---> utils.py           # common utility functions
         |     |---> loader.py          # Core loader for training
         |-> demo.py                    # Demo file
         |-> options.py                 # Set command line options
         |-> train.py                   # Train file
git clone https://github.com/kvmanohar22/DIM.git
cd DIM

Append root of this directory to PYTHONPATH environment variable

export PYTHONPATH=$PYTHONPATH:/path/to/DIM_directory/

Demo

bash ./scripts/demo.sh

The above script gives you a demo of images in img directory.

Full list of options

$ python demo.py --help
usage: demo.py [-h] [--dataset_root DATASET_ROOT] [--log_root LOG_ROOT]
               [--H H] [--W W] [--train_mode] [--max_epochs MAX_EPOCHS]
               [--batch_size BATCH_SIZE] [--base_lr BASE_LR]
               [--gpu_id GPU_ID] [--ckpt_frq CKPT_FRQ] [--alpha ALPHA]
               [--epsilon EPSILON] [--ckpt_path CKPT_PATH]
               [--img_path IMG_PATH] [--tri_path TRI_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --dataset_root DATASET_ROOT
                        Root of dataset directory
  --log_root LOG_ROOT   Directory to log data
  --H H
  --W W
  --train_mode
  --max_epochs MAX_EPOCHS
  --batch_size BATCH_SIZE
  --base_lr BASE_LR
  --gpu_id GPU_ID
  --ckpt_frq CKPT_FRQ   Frequency at which checkpoints are generated
  --alpha ALPHA         Alpha value to weigh two types of losses
  --epsilon EPSILON     Small value used in loss computation
  --ckpt_path CKPT_PATH
                        Path to Checkpoint
  --img_path IMG_PATH   Path to Image
  --tri_path TRI_PATH   Path to Trimap

If you want to test on a custom image, then run the following: bash ./scripts/demo.sh path_to_image path_to_trimap

Train

You can use the following script to train your custom dataset. The data directory should be as follows:

/path/to/data
      |-> input/                # Image directory
      |     |-> img_0.png
      |     |-> img_1.png
      |     |-
      |     |-> img_X.png
      |-> gt/                   # Ground truth alpha mattes
      |     |-> img_0.png
      |     |-> img_1.png
      |     |-
      |     |-> img_X.png
      |-> trimap/               # Trimaps
      |     |-> trimap_dir1/
      |     |     |-> img_0.png
      |     |     |-> img_1.png
      |     |     |-
      |     |     |-> img_X.png
      |     |-> trimap_dir2/
      |     |     |-> img_0.png
      |     |     |-> img_1.png
      |     |     |-
      |     |     |-> img_X.png

bash ./scripts/train.sh path_to_root_of_train_dataset

To set more options, check out the options.py file and set them accordingly in scripts/train.sh

Experiments

Variation of loss with different learning rate schemes

Experimental results

From left to right. Background image, fore-ground image, predicted alpha matte, predicted blended image. Note that image are cropped.

Things which I tried that worked / did not work