Skip to content

Latest commit

 

History

History
45 lines (34 loc) · 1.95 KB

README.md

File metadata and controls

45 lines (34 loc) · 1.95 KB
Raw

Human Segmentation

Description

Deep human segmentation and visual transformation of your photos. Currently supported visual transforms:

  • Bokeh effect
  • Black and white background
  • Layer of mask above person

Examples

Bokeh effect

Black and white background

Layered mask effect

Installation

  • Install requirements: pip3 install -r requirements.txt
  • Download resnet-50 model weights (about 150mb) from google drive
  • Put weights in weights/ folder

Training

Model was trained with supervisely person segmentation dataset. In utils/ folder you can find scripts, that convert supervisely format annotations into binary masks.

You can train any model from qubvel repo. I used Unet architecture with resnet-50 backbone as main model and more lightweight efficientnet-b1.

Notebooks with training code can be found in notebooks/ folder.

Usage

Results available via command line interface. There are some keys:

  • --model_path - path to model weights
  • --device - device to run inference on: gpu or cpu
  • --trans_type - type of visual transformation effect: bokeh, bnw, layered.
  • --result_path - path to save resulting transformed image.
  • --blur_power - only for bokeh transformation option, strength of gaussian blur. Int value from 1 to 3, 1 - the weakest, 3 - the strongest.

Example of console script:

python3 inference.py 'test_photo/medium.jpg' --model_path 'weights/resnet50_089.pb' --device 'gpu' --trans_type 'layered' --result_path 'test1.jpg' --blur_power 2

Citiation

Thanks qubvel for pytorch image segmentation library.