Skip to content

Official repository of Learning Multi-Human Optical Flow (IJCV 2019)

License

Notifications You must be signed in to change notification settings

anuragranj/humanflow2

Repository files navigation

Humanflow2

This is an official repository

Anurag Ranjan, David T. Hoffmann, Dimitrios Tzionas, Siyu Tang, Javier Romero, and Michael J. Black. Learning Multi-Human Optical Flow. IJCV 2019.

[Project Page] [Arxiv]

Prerequisites

Download Multi-Human Optical Flow dataset from here.

Download pre-trained PWC-Net models from NVlabs/PWC-Net and store them in models/ directory.

Install Pytorch. Install dependencies using

pip3 install -r requirements.txt

If there are issues with the correlation module, compile it from source - ClementPinard/Pytorch-Correlation-extension.

Training

For finetuning SPyNet on Multi-Human Optical Flow dataset use:

python main.py PATH_TO_DATASET --dataset humanflow -a spynet --div-flow 1 -b8 -j8 --lr LEARNING_RATE -w 1.0 1.0 1.0 1.0 1.0 --name NAME_OF_EXPERIMENT

For finetuning PWC-Net on Multi-Human Optical Flow dataset use:

python main.py PATH_TO_DATASET --dataset humanflow -a pwc --div-flow 20 -b8 -j8 --lr LEARNING_RATE --name NAME_OF_EXPERIMENT

Testing

To test SPyNet trained on Multi-Human Optical Flow dataset, use

python test_humanflow.py PATH_TO_DATASET --dataset humanflow --arch spynet --div-flow 1 --pretrained pretrained/spynet_MHOF.pth.tar

To test PWC-Net trained on Multi-Human Optical Flow dataset, use

python test_humanflow.py PATH_TO_DATASET --dataset humanflow --arch pwc --div-flow 20 --no-norm  --pretrained pretrained/pwc_MHOF.pth.tar

Acknowledgements

We thank Clement Pinard for his github repository ClementPinard/FlowNetPytorch. We use it as our code base. PWCNet is taken from NVlabs/PWC-Net. SPyNet implementation is taken from sniklaus/pytorch-spynet. Correlation module is taken from ClementPinard/Pytorch-Correlation-extension.

About

Official repository of Learning Multi-Human Optical Flow (IJCV 2019)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published