/deploy_core3d

Dockerized Modules for Shadow-Removal, RGB/RGBD Inpainting, MSI2RGB and Joint Semantic Segmentation/ Single-View DHM estimation on satellite imagery

Primary LanguagePythonOtherNOASSERTION

Modules for Shadow-Removal, Inpainting, MSI2RGB and Joint Semantic Segmentation/ Single-View DHM estimation

placeholder The list of modules are as follows:

  • Joint Shadow Removal and Shadow Probability Estimation for Satellite Imagery.
  • Inpainting
    1. RGB to RGB inpainting
    2. RGBD to RGBD tiled iterative inpainting.
  • 8-band MSI (unknown normalization/sensor data) to photo-realistic RGB, shadow-free RGB and shadow probabilities.
  • Joint Semantic Segmentation and single-view DHM estimation from satellite imagery.

Usage (using pre-built docker images on dockerhub)

Each module has its own docker image that is built using NVcaffe-v17, cuda 10.1 and cudnn-v7 on ubuntu16.04:

To run any module, do docker run --gpus 1 -it venkai/[module-name] and follow the instructions in the prompt.

Note that you need a CUDA capable GPU and NVIDIA Linux x86_64 Driver Version >= 418.39.

Custom Build Instructions (optional)

If you wish to build NVCaffe from scratch, navigate to ./nvcaffe_docker and build the Dockerfile using docker build -t venkai/nvcaffe . If you have an older NVIDIA driver that you don't want to upgrade, you can modify the first line of the Dockerfile with the appropriate nvidia/cuda image using the compatibility table provided here.

To build any module, a generic Dockerfile template is provided in ./software/Dockerfile.generic. Navigate to any module sub-folder and build the docker image using docker build -t [software-name] -f [path to Dockerfile.generic]. Example: navigate to software/shadow_removal and build using docker build -t venkai/shadow-removal -f ../Dockerfile.template

License & Citation

All modules are released under a variant of the BSD 2-Clause license.

If you find any of our modules useful in your research, please consider citing our relevant papers:

@inproceedings{santhanam2017generalized,
    title={Generalized deep image to image regression},
    author={Santhanam, Venkataraman and Morariu, Vlad I and Davis, Larry S},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    pages={5609--5619},
    year={2017}
}

@ARTICLE{8804390,
    author={V. {Santhanam} and L. S. {Davis}},
    journal={IEEE Transactions on Neural Networks and Learning Systems},
    title={A Generic Improvement to Deep Residual Networks Based on Gradient Flow},
    year={2019}, volume={}, number={}, pages={1-10},
}

Acknowledgements

The research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via DOI/IBC Contract Number D17PC00287. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.