Skip to content

Latest commit

 

History

History
128 lines (112 loc) · 5.16 KB

DATASET.md

File metadata and controls

128 lines (112 loc) · 5.16 KB

DATASET

Download available datasets

The dataset is available on Hugging Face Datasets. Please download the set you need:

Dataset Image/video Train/val Filenames
I-HIM50K Image Train I-HIM50K.z* (6 files)
M-HIM2K + HIM2K Image Val HIM2K_M-HIM2K.zip
V-HIM2K5 Video Train V-HIM2K5.z* (7 files)
V-HIM60 Video Val V-HIM60.zip

For the zip splits, you can run following command to combine before unzip:

# e.g, if you want to combine I-HIM50K files
zip -F I-HIM50K.zip --out I-HIM50K_all.zip

# then unzip the file
unzip I-HIM50K_all.zip

How to synthesize dataset?

I-HIM50K

You need to download the HHM dataset, BG-20K dataset and invalid_him.txt before running this script:

python tools/synthesize_image_him.py --image-root HHM/train/images \
                                     --invalid-names invalid_him.txt \
                                     --bg-root BG-20K/train \
                                     --output-dir data/I-HIM50K \
                                     --n-workers 80

Here, we run the process in 80 threads, please adjust it based on your machine.

M-HIM2K

To gen all binary masks for HIM2K, we use detectron2. We provide scripts to process masks in tools/gen_mask. The process of generate masks are:

  1. Clone the detectron2 repository.
  2. Copy all files in tools/gen_mask to demo directory.
  3. Download all pretrained models from Model Zoo to detectron2/pretrained
  4. Update data paths in get_mask_single.sh
  5. Generate masks with gen_mask_all.sh

V-HIM

You need to download those datasets

  1. VideoMatte240K
  2. VM108
  3. CRGNN

and foreground/background list:

  1. fg_train.txt
  2. fg_test.txt
  3. bg_train.txt
  4. bg_test.txt

After organizing the data following paths in *.txt files, you can generate the V-HIM2K5 by running this command:

python tools/synthesize_video_him.py --split train --data-dir /path/to/video --out-dir data/V-HIM2K5 --n-workers 80
# or
python tools/synthesize_video_him.py --split test --data-dir /path/to/video --out-dir data/V-HIM60 --n-workers 16

To generate masks, we combine the ouput of GenVIS and XMem. The complete code will be released in the future!

How to use your own dataset?

You should prepare your data following these structures

For image matting training dataset

<root_dir>
└── <split>
    ├── images
    │   ├── image1.jpg
    │   └── image2.jpg
    ├── <alpha_dir_name>
    │   ├── image1
    │   │   ├── 00.png # instance matte 1
    │   │   └── 01.png # instance matte 2
    │   └── image2
    └──<mask_dir_name (optional)>
        ├── image1
        │   ├── 00.png # instance mask 1
        │   └── 01.png # instance mask 2
        └── image2

For image matting validation/testing dataset

<root_dir>
├── images
|   └── <split>
│       ├── image1.jpg
│       └── image2.jpg
├── <alpha_dir_name>
|   └── <split>
│       ├── image1
│       │   ├── 00.png # instance matte 1
│       │   └── 01.png # instance matte 2
│       └── image2
└──<mask_dir_name>
    └── <split>
        ├── image1
        │   ├── 00.png # instance mask 1
        │   └── 01.png # instance mask 2
        └── image2

For video matting dataset

<root_dir>
└── <split>
    ├── fgr
    │   └── video1
    │       ├── image1.jpg
    │       └── image2.jpg
    ├── <alpha_dir_name>
    │   └── video1
    │       ├── image1
    │       │   ├── 00.png # instance matte 1
    │       │   └── 01.png # instance matte 2
    │       └── image2
    └──<mask_dir_name (optional in training)>
        └── video1
            ├── image1
            │   ├── 00.png # instance mask 1
            │   └── 01.png # instance mask 2
            └── image2