Pytorch_CamouflageImages

Generate Camouflage Images by Pytorch

this implementation is mainly based on Deep Camouflage Images

Usage

Setup Environment

Dependencies

• PyTorch (>= 1.7)
• torchvision
• tqdm
• albumentations
• scikit-learn==0.23.2
• opencv-contrib

In the case of conda.

conda create -n camouflage python=3.8
conda activate camouflage

and then, install pytorch / torchvision, reference to official website

finally, install dependent libraries

pip install -r requirements.txt

Test

python camouflage_HRNet.py --params <path to parameter python file>

For quick test, you can generate camouflage images by python camouflage_HRNet.py --params params_Canyon

Here is process window

|

Preparaion

For custom images, you should prepare

• background image (ex. samples/inputs/cliff.jpg)
• foreground image (ex. samples/inputs/kuma.png)
• foreground image mask (ex.samples/inputs/kuma_mask/png)
  • the size of foreground image mask must be same as that of foreground image !
• .py file with parameters corresponding for your data (ex. params_Cliff.py)

Parameters

initial setting

- input_path : path to foreground image
- mask_path : path to foreground image mask
- bg_path : path to background image
- output_dir : directory for generated images
- name : prefix name for generated image
- seed : pytorch seed

mask setting

- mask_scale : scale ratio for foreground image mask
- crop : crop foreground image mask to bounding box
- hidden_selected : If None, use hidden recommendation. If you wouldn't like to use hidden recommendation, you should give [y1_start,x1_start] for this parameter

train setting

- epoch : iteration for training
- lr : learning rate for adam

loss setting

- erode_border : If True, erode attention map
- style_weight_dic : dictionary of weights for style loss
- style_all : If True, use all background image for style loss. If False, use corresponding background image for style loss.
- alpha1 : scale parameter for leave loss
- alpha2 : scale parameter for remove loss
- mu : ratio of remove loss for camouflage loss
- lambda_weights : dictionary of weights for all loss

log setting

- show_every : interval for displaying intermediate result
- save_process : If True, save intermediate result by every `show_every`
- show_comp : compressino ratio for display

Influence by loss function

According to Deep Camouflage Images, losses has following impact for generated image:

• style loss : control similarity between generated image and background image
• camouflage loss : control diffuculty for detection of camouflage objects in generated image
  • leave loss : leave foreground features in generated image
  • remove loss : remove foreground features in generated image
• reguralization loss : control consistency for generated image
• total variation loss : smooth generated image

style	style+cam

style+cam+reg	style+cam+reg+tv

Gallary

SeaMountain	CliffRiver

Mountain	Cliff