Skip to content
master
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
doc
 
 
 
 
 
 
 
 
 
 

README.md



CompenNet++: End-to-end Full Projector Compensation (ICCV'19)

result1

Introduction

PyTorch implementation of CompenNet++. Also see journal version.

Highlights:

  • The proposed CompenNet++ is the first end-to-end full projector compensation system.
  • Compared with two-step methods (e.g., CompenNet w/ SL), CompenNet++ learns the geometric correction without extra sampling images (~42 images) and outperforms the compared counterparts.
  • Two task-specific weight initialization approaches are proposed to ensure the convergence and stability of CompenNet++.
  • Novel simplification techniques are developed to improve the running time efficiency of CompenNet++.

For more info please refer to our ICCV'19 paper, high-res supplementary material ~(180M) and CompenNet++ benchmark dataset (~11G).

Prerequisites

  • PyTorch compatible GPU
  • Python 3
  • PyTorch >= 0.4.0
  • opencv-python 3.4.4
  • visdom (for visualization)

Usage

  1. Clone this repo:

     git clone https://github.com/BingyaoHuang/CompenNet-plusplus
     cd CompenNet-plusplus
    
  2. Install required packages by typing

     pip install -r requirements.txt
    
  3. Download CompenNet++ benchmark dataset (~11G) and extract to data/

  4. Start visdom by typing

     visdom
    
  5. Once visdom is successfully started, visit http://localhost:8097 (train locally) or http://serverhost:8097 (train remotely).

  6. Open main.py and set which GPUs to use. An example is shown below, we use GPU 0, 2 and 3 to train the model.

     os.environ['CUDA_VISIBLE_DEVICES'] = '0, 2, 3'
     device_ids = [0, 1, 2]
    
  7. Run main.py to start training and testing

     cd src/python
     python main.py
    
  8. The training and validation results are updated in the browser during training. An example is shown below, where the 1st figure shows the training and validation loss, rmse and ssim curves. The 2nd and 3rd montage figures are the training and validation pictures, respectively. In each montage figure, the 1st rows are the camera captured uncompensated images, the 2nd rows are CompenNet++ predicted projector input images and the 3rd rows are ground truth of projector input images.

  9. The quantitative comparison results will be saved to log/%Y-%m-%d_%H_%M_%S.txt after training.

visdom


Apply CompenNet++ to your own setup

  1. For a nonplanar textured projection surface, adjust the camera-projector such that the brightest projected input image (plain white data/ref/img_0125.png) slightly overexposes the camera captured image. Similarly, the darkest projected input image (plain black data/ref/img_0001.png) slightly underexposes the camera captured image. This allows the projector dynamic range to cover the full camera dynamic range.
  2. Once the setup is fixed, we create a setup data directory data/light[n]/pos[m]/[surface] (we refer it to data_root), where [n] and [m] are lighting and pose setup indices, respectively. [surface] is the projection surface's texture name.
  3. Project and capture the plain black data/ref/img_0001.png and the plain white images data/ref/img_0125.png for projector FOV mask detection later. Then, save the captured images to data_root/cam/raw/ref/img_0001.png(img_0125.png).
  4. Project and capture a surface image data/ref/img_gray.png. Then, save the captured images to data_root/cam/raw/ref/img_0126.png.
  5. Project and capture the training and validation images in data/train and /data/test. Then, save the captured images to data_root/cam/raw/train, data_root/cam/raw/test, respectively.
  6. Find the optimal displayable area following the algorithm in loadData in trainNetwork.py. Then, affine transform the images in data/test to the optimal displayable area and save transformed images to data_root/cam/raw/desire/test. Refer to model testing below.

Note other than ref/img_0001.png, ref/img_0125.png and ref/img_gray.png, the rest plain color images are used by original TPS w/ SL method, we don't need them to train CompenNet++. Similarly, data_root/cam/raw/sl and data_root/cam/warpSL are only used by two-step methods.


Network architecture (training)

train

Network architecture (testing)

test

Citation

@inproceedings{huang2019compennet++,
    author = {Huang, Bingyao and Ling, Haibin},
    title = {CompenNet++: End-to-end Full Projector Compensation},
    booktitle = {IEEE International Conference on Computer Vision (ICCV)},
    month = {October},
    year = {2019} }

@inproceedings{huang2019compennet,
    author = {Huang, Bingyao and Ling, Haibin},
    title = {End-To-End Projector Photometric Compensation},
    booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2019} }

Acknowledgments

The PyTorch implementation of SSIM loss is modified from Po-Hsun-Su/pytorch-ssim. The PyTorch implementation of TPS warping is modified from cheind/py-thin-plate-spline. We thank the anonymous reviewers for valuable and inspiring comments and suggestions. We thank the authors of the colorful textured sampling images.

License

This software is freely available for non-profit non-commercial use, and may be redistributed under the conditions in license.

About

[ICCV 2019] CompenNet++: End-to-end Full Projector Compensation

Topics

Resources

License

Releases

No releases published

Packages

No packages published

Languages

You can’t perform that action at this time.