Automatic Image Background Subtraction

This repo contains set of scripts for automatic one-shot image background subtraction task using the following strategies:

1. the appropriate background subtraction services (mostly human based):

• benzin.io,
• remove.bg.

2. U-Net human segmentation net + CascadePSP refinement net:
3. BackgroundMattingV2 net.

Installation

git clone https://github.com/osmr/bgsub.git
cd bgsub
pip install -r requirements.txt

Usage

1. Launch a script for background subtraction via benzin.io/remove.bg service:

python subtract_bg_service.py --service=<service> --token=<your token> --input=<directory with images> --output=<output directory with binary masks>

Here:

• service is benzinio for benzin.io service or removebg for remove.bg,
• token is a service API token value, which you will receive after registering on the selected service,
• input is a directory with processing JPEG images (can contain subdirectories),
• output is a directory with resulted PNG binary masks (it is assumed that all original images had unique names).

Optional parameters:

• middle is a directory with intermediate images with original masks obtained from the service (PNG with alpha mask),
• ppdir is a flag for adding extra parrent+parrent directory to the output one (should use as --ppdir).
• threshold is a threshold for mask binarization (default value is 127),
• url is an optional custom URL for service,
• jpg is a flag for forced recompression an input image as JPG (should use as --jpg),
• not-resize is a flag for suppressing forcible scale the mask to the input image (should use as --not-resize).

2. Launch a script for background subtraction via human segmentation network:

python subtract_bg_human.py --input=<directory with images> --output=<output directory with binary masks>

Here:

• input is a directory with processing JPEG images (can contain subdirectories),
• output is a directory with resulted PNG binary masks (it is assumed that all original images had unique names).

Optional parameters:

• ppdir is a flag for adding extra parrent+parrent directory to the output one (should use as --ppdir).
• use-cuda is a flag for using CUDA for network inference (should use as --use-cuda).

3. Launch a script for background subtraction via matting network:

python subtract_bg_matting.py --input=<directory with images> --bg=<background image path> --output=<output directory with binary masks>

Here:

• input is a directory with processing JPEG images (can contain subdirectories),
• bg is a background image file path,
• output is a directory with resulted PNG binary masks (it is assumed that all original images had unique names).

Optional parameters:

• threshold is a threshold for mask binarization (default value is 127),
• ppdir is a flag for adding extra parrent+parrent directory to the output one (should use as --ppdir).
• use-cuda is a flag for using CUDA for network inference (should use as --use-cuda).

Remark

The script does not recalculate the masks if the target images already exist.