flownet2-pytorch

Pytorch implementation of FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks.

Multiple GPU training is supported, and the code provides examples for training or inference on MPI-Sintel clean and final datasets. The same commands can be used for training or inference with other datasets. See below for more detail.

Inference using fp16 (half-precision) is also supported.

For more help, type

python main.py --help

Network architectures

Below are the different flownet neural network architectures that are provided.
A batchnorm version for each network is available.

FlowNet2S
FlowNet2C
FlowNet2CS
FlowNet2CSS
FlowNet2SD
FlowNet2

Custom layers

FlowNet2 or FlowNet2C* achitectures rely on custom layers Resample2d or Correlation.
A pytorch implementation of these layers with cuda kernels are available at ./networks.
Note : Currently, half precision kernels are not available for these layers.

Data Loaders

Dataloaders for FlyingChairs, FlyingThings, ChairsSDHom and ImagesFromFolder are available in datasets.py.

Loss Functions

L1 and L2 losses with multi-scale support are available in losses.py.

Installation

# get flownet2-pytorch source
git clone https://github.com/NVIDIA/flownet2-pytorch.git
cd flownet2-pytorch

# install custom layers
bash install.sh

Docker image

Libraries and other dependencies for this project include: Ubuntu 16.04, Python 2.7, Pytorch 0.2, CUDNN 6.0, CUDA 8.0

A Dockerfile with the above dependencies is available

# Build and launch docker image
bash launch_docker.sh

Converted Caffe Pre-trained Models

We've included caffe pre-trained models. Should you use these pre-trained weights, please adhere to the license agreements.

FlowNet2[620MB]
FlowNet2-C[149MB]
FlowNet2-CS[297MB]
FlowNet2-CSS[445MB]
FlowNet2-CSS-ft-sd[445MB]
FlowNet2-S[148MB]
FlowNet2-SD[173MB]

Inference

# Example on MPISintel Clean   
python main.py --inference --model FlowNet2 --save_flow --inference_dataset MpiSintelClean \
--inference_dataset_root /path/to/mpi-sintel/clean/dataset \
--resume /path/to/checkpoints

Training and validation

# Example on MPISintel Final and Clean, with L1Loss on FlowNet2 model
python main.py --batch_size 8 --model FlowNet2 --loss=L1Loss --optimizer=Adam --optimizer_lr=1e-4 \
--training_dataset MpiSintelFinal --training_dataset_root /path/to/mpi-sintel/final/dataset  \
--validation_dataset MpiSintelClean --validation_dataset_root /path/to/mpi-sintel/clean/dataset

# Example on MPISintel Final and Clean, with MultiScale loss on FlowNet2C model 
python main.py --batch_size 8 --model FlowNet2C --optimizer=Adam --optimizer_lr=1e-4 --loss=MultiScale --loss_norm=L1 \
--loss_numScales=5 --loss_startScale=4 --optimizer_lr=1e-4 --crop_size 384 512 \
--training_dataset FlyingChairs --training_dataset_root /path/to/flying-chairs/dataset  \
--validation_dataset MpiSintelClean --validation_dataset_root /path/to/mpi-sintel/clean/dataset

Results on MPI-Sintel

Reference

If you find this implementation useful in your work, please acknowledge it appropriately and cite the paper using:

@InProceedings{IMKDB17,
  author       = "E. Ilg and N. Mayer and T. Saikia and M. Keuper and A. Dosovitskiy and T. Brox",
  title        = "FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks",
  booktitle    = "IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
  month        = "Jul",
  year         = "2017",
  url          = "http://lmb.informatik.uni-freiburg.de//Publications/2017/IMKDB17"
}

Acknowledgments

Parts of this code were derived, as noted in the code, from ClementPinard/FlowNetPytorch.

skic/flownet2-pytorch