End-to-end Full Projector Compensation (PAMI)

Introduction

PyTorch implementation of CompenNeSt++ (PAMI), a journal extension of CompenNet (CVPR'19) and CompenNet++ (ICCV'19)

Highlights:

• The proposed CompenNeSt++ is the first end-to-end full projector compensation system.
• CompenNeSt++ significantly outperforms CompenNet++ (ICCV'19) and has 27.7% less parameters.
• We propose a pre-trained CompenNeSt++ that can work with only 8 images and 3 min training time (see CompeNeSt++ pre in the figure above).
• We perform in-depth ablation studies and show what features are learned by CompenNeSt++ and how the compensation is performed in the feature space. Such studies were not available in our previous papers.
• Compared with two-step methods, CompenNeSt++ not only is fully differentiable but also learns the geometric correction without an additional step of structured light and outperforms the traditional two-step methods (e.g., TPS w/ SL).
• Two task-specific weight initializations and two network simplification techniques are proposed to further improve the convergence and running time efficiency of CompenNeSt++.
• For the first time, a synthetic data-based pre-training method is proposed to significantly improve the practical efficiency of our system.
• For the first time, a setup-independent full compensation benchmark and a partial photometric compensation benchmark are constructed.

For more info please refer to our PAMI paper and supplementary material (~90M).

---

Learning-based formulation

We disentangle the geometry and photometry from the complex projector-camera light transport and model them as WarpingNet and CompenNeSt, respectively. Note that the blue and red paths share the same physical process, thus we can learn CompenNeSt++ from the backward mapping \tilde{x} -> x (i.e., reverse blue path) instead of \tilde{x*} -> x* (i.e., reverse red path), obviating the need for the ground truth compensation image x*. See CompenNeSt++ Sec. 3.1-3.2 and Fig. 3 for more details.

---

Network architecture (training)

CompenNeSt++ inferred projector input as training progresses

Network architecture (testing)

---

Prerequisites

• PyTorch compatible GPU
• Python 3
• PyTorch >= 1.2.0
• opencv-python 3.4.4
• visdom (for visualization)

Benchmark dataset

• Full compensation (~11GB, 20 setups, same as CompenNet++ dataset, but add cam\warpSL\desire\test)
• (optional) Photometric compensation (~4G, 24 setups, same as CompenNet dataset, images are geometrically registered)
• (optional) Blender synthetic photometric compensation (~40GB, 100 setups, we use it for CompenNeSt pre-training)

Usage

1. Clone this repo:

    git clone https://github.com/BingyaoHuang/CompenNeSt-plusplus
    cd CompenNeSt-plusplus
   

2. Install required packages by typing

    pip install -r requirements.txt
   

3. Download CompenNeSt++ full compensation dataset 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 train_compenNeSt++.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 train_compenNeSt++.py to start training and testing

    cd src/python
    python train_compenNeSt++.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 (\tilde{x}), the 2nd rows are CompenNeSt++ inferred projector input images (\hat{x}) and the 3rd rows are ground truth of projector input images (x).

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

---

Apply CompenNeSt++ 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. 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.
7. Update data_list in train_compenNeSt++.py (or train_pre-trained_compenNeSt++.py that only requires 8 training images) and run the script.
8. Project the compensated projector input images under data_root/prj/cmp/test/[model].

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

---

Citation

@article{huang2020CompenNeSt++,
    title={End-to-end Full Projector Compensation},
    author={Bingyao Huang and Tao Sun and Haibin Ling},
    year={2021},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)} }

@inproceedings{huang2019CompenNet++,
    author = {Huang, Bingyao and Ling, Haibin},
    title = {CompenNeSt++: 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.