Awesome

liteflownet-tf2

LiteFlowNet inference realization with TensorFlow 2.

Inspired by pytorch reimplemetation: https://github.com/sniklaus/pytorch-liteflownet

This is my reimplementation of LiteFlowNet [1] using Tensorflow 2.0 with build-in correlation layer. There are some messy code with using of tf2 package and it compatibility mode with tf 1.x. Also, please cite the paper accordingly and make sure to adhere to the licensing terms of the authors. Should you be making use of this particular implementation, please acknowledge it appropriately [3].

Original Caffe version: https://github.com/twhui/LiteFlowNet

Setup

The correlation layer is using tensorflow_addons package which required tf2+ version. These repo tested with packages installed via pip:

pip install --user tensorflow==2.0.0
pip install --user tensorflow_addons==0.6.0

Weights could be converted from PyTorch version [2] via script convert_pytorch2tf.py and directly from Caffe version [1].

I converted only default model. You can convert it as youself. Or you can download already converted model.

Convert weights from pytorch version:

# download model
wget http://content.sniklaus.com/github/pytorch-liteflownet/network-default.pytorch

# convert default model
python convert_pytorch2tf.py --input_model=./network-default.pytorch --output_model=./model

Convert weights from caffe version:

# download and unzip model
wget https://github.com/twhui/LiteFlowNet/raw/master/models/trained/liteflownet.tar.gz
tar -xvzf liteflownet.tar.gz

# convert default model
python convert_caffe2tf.py --input_model=./liteflownet.caffemodel --output_model=./model

Get converted weights to tf2

You can download already converted model from the link: https://drive.google.com/open?id=1apeRotQKMsFji8MKKNzcx-QO4udkJcdJ.

How to run

To run it on your own pair of images, use the following command.

python eval.py --img1=./images/first.png --img2=./images/second.png --flow=./out.flow --display_flow=True

Results differ a little. I think it depends on a bit different feature warping then in original work.

License

Original materials is provided for research purposes only. Please see https://github.com/twhui/LiteFlowNet#license-and-citation to more information.

References

[1]  @inproceedings{Hui_CVPR_2018,
         author = {Tak-Wai Hui and Xiaoou Tang and Chen Change Loy},
         title = {{LiteFlowNet}: A Lightweight Convolutional Neural Network for Optical Flow Estimation},
         booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
         year = {2018}
     }

[2]  @misc{pytorch-liteflownet,
         author = {Simon Niklaus},
         title = {A Reimplementation of {LiteFlowNet} Using {PyTorch}},
         year = {2019},
         howpublished = {\url{https://github.com/sniklaus/pytorch-liteflownet}}
    }

[2]  @misc{liteflownet-tf2,
         author = {Vladimir Mikhelev},
         title = {{LiteFlowNet} inference realization with {Tensorflow 2}},
         year = {2020},
         howpublished = {\url{https://github.com/keeper121/liteflownet-tf2}}
    }