wuhuikai / DeepGuidedFilter

Official Implementation of Fast End-to-End Trainable Guided Filter, CVPR 2018

Fast End-to-End Trainable Guided Filter

[Project] [Paper] [arXiv] [Demo] [Home]

Official implementation of Fast End-to-End Trainable Guided Filter.
Faster, Better and Lighter for pixel-wise image prediction.

Overview

DeepGuidedFilter is the author's implementation of:

Fast End-to-End Trainable Guided Filter
Huikai Wu, Shuai Zheng, Junge Zhang, Kaiqi Huang
CVPR 2018

With our method, FCNs can run 10-100 times faster w/o performance drop.

Contact: Hui-Kai Wu ([e-mail])

Get Started

Prepare Environment [Python>=3.6]

1. Download source code from GitHub.

   git clone https://github.com/wuhuikai/DeepGuidedFilter
   
   cd DeepGuidedFilter && git checkout release

2. Install dependencies.

   conda install opencv=3.4
   conda install pytorch=1.1 torchvision=0.2 cudatoolkit=9.0 -c pytorch
   
   pip install -r requirements.txt 

3. (Optional) Install dependencies for MonoDepth.

   cd ComputerVision/MonoDepth
   
   pip install -r requirements.txt

Ready to GO !

Image Processing

cd ImageProcessing/DeepGuidedFilteringNetwork

python predict.py  --task auto_ps \
                   --img_path ../../images/auto_ps.jpg \
                   --save_folder . \
                   --model deep_guided_filter_advanced \
                   --low_size 64 \
                   --gpu 0

See Here or python predict.py -h for more details.

Semantic Segmentation with Deeplab-Resnet

1. Enter the directory.

   cd ComputerVision/Deeplab-Resnet

2. Download the pretrained model [Google Drive|BaiduYunPan].
3. Run it now !

   python predict_dgf.py --img_path ../../images/segmentation.jpg --snapshots [MODEL_PATH]

Note:

1. Result is in ../../images.
2. Run python predict_dgf.py -h for more details.

Saliency Detection with DSS

1. Enter the directory.

   cd ComputerVision/Saliency_DSS

2. Download the pretrained model [Google Drive|BaiduYunPan].
3. Try it now !

   python predict.py --im_path ../../images/saliency.jpg \
                     --netG [MODEL_PATH] \
                     --thres 161 \
                     --dgf --nn_dgf \
                     --post_sigmoid --cuda

Note:

1. Result is in ../../images.
2. See Here or python predict.py -h for more details.

Monocular Depth Estimation

1. Enter the directory.

   cd ComputerVision/MonoDepth

2. Download and Unzip Pretrained Model [Google Drive|BaiduYunPan]
3. Run on an Image

   python monodepth_simple.py --image_path ../../images/depth.jpg --checkpoint_path [MODEL_PATH] --guided_filter

Note:

1. Result is in ../../images.
2. See Here or python monodepth_simple.py -h for more details.

Guided Filtering Layer

Install Released Version

• PyTorch Version

  pip install guided-filter-pytorch

• Tensorflow Version

  pip install guided-filter-tf

Usage

• PyTorch Version

  from guided_filter_pytorch.guided_filter import FastGuidedFilter
  
  hr_y = FastGuidedFilter(r, eps)(lr_x, lr_y, hr_x)

  from guided_filter_pytorch.guided_filter import GuidedFilter
  
  hr_y = GuidedFilter(r, eps)(hr_x, init_hr_y)

  from guided_filter_pytorch.guided_filter import ConvGuidedFilter
  
  hr_y = ConvGuidedFilter(r, norm)(lr_x, lr_y, hr_x)
  

• Tensorflow Version

  from guided_filter_tf.guided_filter import fast_guided_filter
  
  hr_y = fast_guided_filter(lr_x, lr_y, hr_x, r, eps, nhwc)

  from guided_filter_tf.guided_filter import guided_filter
  
  hr_y = guided_filter(hr_x, init_hr_y, r, eps, nhwc)

Training from scratch

Prepare Training Environment

git checkout master

conda install opencv=3.4
conda install pytorch=1.1 torchvision=0.2 cudatoolkit=9.0 -c pytorch

pip uninstall Pillow
pip install -r requirements.txt

# (Optional) For MonoDepth
pip install -r ComputerVision/MonoDepth/requirements.txt 

Start to Train

• Image Processing
• Semantic Segmentation with Deeplab-Resnet
• Saliency Detection with DSS
• Monocular Depth Estimation

Citation

@inproceedings{wu2017fast,
  title     = {Fast End-to-End Trainable Guided Filter},
  author    = {Wu, Huikai and Zheng, Shuai and Zhang, Junge and Huang, Kaiqi},
  booktitle = {CVPR},
  year = {2018}
}