491913145/optical_flow_net-PWC-Net_python3_PyTorch

optical_flow_net-PWC-Net_python3_PyTorch

This is the implementation of the "FlowNet: Learning Optical Flow with Convolutional Networks" paper. It constructs appropriate CNNs which are capable of solving the optical flow estimation problem using Python 3.6, PyTorch and runs on the GPU of colab.

Installation

The code was developed using Python 3.6 & PyTorch 1.4.0 & CUDA 10.1, using colab.

Test

• Test the code: execute Optical flow python3 pytorch.ipynb. You can set your input images by setting the variables im1_fn and im2_fn. You can set the directory of the .flo output by setting the variable flow_fn.
• pwc_net.pth.tar is the fine-tuned weight on MPI Sintel

Running the tests

• Input 1 Image1
• Input 2 Image2
• Ground Truth ground_truth
• Output .flo file frame_output.flo
• Output .png image frame_output.png

Acknowledgments

• Thanks to Dr. Marwan Torki for explaining to me how the correlation works

Contact

Aya Lotfy (ayalotfy2019@gmail.com)

Related Work

• PWC-Net
• PReMVOS

Paper & Citation

@inproceedings{DFIB15,
  author       = "A. Dosovitskiy and P. Fischer and E. Ilg and P. H{\"a}usser and C. Haz\ırba\ş and V. Golkov and P. v.d. Smagt and D. Cremers and T. Brox",
  title        = "FlowNet: Learning Optical Flow with Convolutional Networks",
  booktitle    = "IEEE International Conference on Computer Vision (ICCV)",
  month        = "Dec",
  year         = "2015",
  url          = "http://lmb.informatik.uni-freiburg.de/Publications/2015/DFIB15"
}