UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss
This repository contains the TensorFlow implementation of the paper
UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss (AAAI 2018)
Simon Meister, Junhwa Hur, and Stefan Roth.
Slides
Download slides from AAAI 2018 talk.
Citation
If you find UnFlow useful in your research, please consider citing:
@inproceedings{Meister:2018:UUL,
title = {{UnFlow}: Unsupervised Learning of Optical Flow
with a Bidirectional Census Loss},
author = {Simon Meister and Junhwa Hur and Stefan Roth},
address = {New Orleans, Louisiana},
booktitle = {AAAI},
month = feb,
year = {2018}
}
License
UnFlow is released under the MIT License (refer to the LICENSE file for details).
Unofficial PyTorch code
There is a unofficial third party PyTorch implementation by Simon Niklaus.
Contents
Introduction
Our paper describes a method to train end-to-end deep networks for dense optical flow without the need for ground truth optical flow.
NOTE (January 2020): There may be some hiccups with more recent versions of tensorflow due to the unstable custom op compilation API used for the correlation operation. If you experience these issues, please try one of the older tensorflow versions (see list of releases, e.g. 1.2 or 1.7) until I find time to fix these issues. I'm currently too busy with new projects to upgrade the code, and would be happy about any contributions.
This implementation supports all training and evaluation styles described in the paper. This includes
- unsupervised training with our proxy loss(es) on SYNTHIA, KITTI raw, Cityscapes, and FlyingChairs,
- supervised fine-tuning on KITTI 2012 flow and KITTI 2015 flow,
- evaluation on KITTI 2012 flow, KITTI 2015 flow, Middlebury flow, and MPI Sintel flow.
The supported network architectures are FlowNetS, FlowNetC, as well as stacked variants of these networks as introduced in FlowNet 2.0.
All datasets required for training and evaluation are downloaded on-demand by the code (except Cityscapes, which has to be downloaded manually if needed). Please ensure that the data directory specified in the configuration file has enough space to hold at least 150 GB (when using SYNTHIA and KITTI only).
Usage
Please refer to the configuration file template (config_template/config.ini) for a detailed description of the different operating modes.
Hardware requirements
- at least one NVIDIA GPU (multi-GPU training is supported). We used the Titan X Pascal with 12GB memory.
- for best performance, at least 8GB of GPU memory is recommended, for the stacked variants 11-12GB may be best
- to run the code with less GPU memory, take a look at https://github.com/openai/gradient-checkpointing. One user reported successfully running the full CSS model with a GTX 960 with 4GB memory. Note that training will take longer in that case.
Software requirements
- python 3
- gcc4
- RAR backend tool for
rarfile(see https://pypi.python.org/pypi/rarfile/) - python packages:
matplotlib pypng rarfile pillowandtensorflow-gpu(at least version 1.7) - CUDA and CuDNN. You should use the installer downloaded from the NVIDIA website and install to /usr/local/cuda. Please make sure that the versions you install are compatible with the version of
tensorflow-gpuyou are using (see e.g. https://github.com/tensorflow/tensorflow/releases).
Prepare environment
- copy
config_template/config.inito./and modify settings in the[dir],[run]and[compile]sections for your environment (see comments in the file).
Run & evaluate experiments
- adapt settings in
./config.inifor your experiment cd src- train with
python run.py --ex my_experiment. Evaluation is run during training as specified inconfig.ini, but no visualizations are saved. - evaluate (multiple) experiments with
python eval_gui.py --ex experiment_name_1[, experiment_name_2 [...]]and visually compare results side-by-side.
You can use the --help flag with run.py or eval_gui.py to view all available flags.
View tensorboard logs
- view logs for all experiments with
tensorboard --logdir=<log_dir>/ex
Pre-trained models
We provide checkpoints for the C_SYNTHIA, CS_SYNTHIA, CSS_SYNTHIA, C, CS, CSS and CSS_ft models (see "Replicating our models" for a description). To use them,
- download this file and extract the contents to
<log_dir>/ex/.
Now, you can evaluate and compare different models, e.g.
python eval_gui.py --ex C_SYNTHIA,C,CSS_ft.
Replicating our models
In the following, each list item gives an experiment name and parameters to set in ./config.ini. For each experiment, first modify the configuration parameters as specified, and then run python run.py --ex experiment_name.
First, create a series of experiments for the models pre-trained on SYNTHIA:
- C_SYNTHIA:
dataset = synthia,flownet = C, - CS_SYNTHIA:
dataset = synthia,flownet = CS,finetune = C_SYNTHIA, - CSS_SYNTHIA:
dataset = synthia,flownet = CSS,finetune = C_SYNTHIA,CS_SYNTHIA.
Next, create a series of experiments for the models trained on KITTI raw:
- C:
dataset = kitti,flownet = C,finetune = C_SYNTHIA, - CS:
dataset = kitti,flownet = CS,finetune = C,CS_SYNTHIA, - CSS:
dataset = kitti,flownet = CSS,finetune = C,CS,CSS_SYNTHIA.
Then, train the final fine-tuned model on KITTI 2012 / 2015:
- CSS_ft:
dataset = kitti_ft,flownet = CSS,finetune = C,CS,CSS,train_all = True. Please monitor the eval logs and stop the training as soon as the validation error starts to increase (for CSS, it should be at about 70K iterations).
If you want to train on Cityscapes:
- C_Cityscapes:
dataset = cityscapes,flownet = C,finetune = C_SYNTHIA.
Note that all models but CSS_ft were trained without ground truth optical flow, using our unsupervised proxy loss only.
Navigating the code
The core implementation of our method resides in src/e2eflow/core. The key files are:
- losses.py: Proxy losses for unsupervised training,
- flownet.py: FlowNet architectures with support for stacking,
- image_warp.py: Backward-warping via differentiable bilinear image sampling,
- supervised.py: Supervised loss and network computation,
- unsupervised.py: Unsupervised (bi-directional) multi-resolution network and loss computation,
- train.py: Implementation of training with online evaluation,
- augment.py: Geometric and photometric data augmentations.
15 Nov 04, 2022
25 Aug 12, 2021
79 Jan 06, 2023
43 Dec 26, 2022
1.8k Jan 03, 2023
69 Dec 21, 2022
47 Nov 15, 2022
10 Oct 24, 2022
19 Jun 07, 2022
29 Sep 24, 2022
5.1k Jan 03, 2023
72 Dec 14, 2022
15 Feb 27, 2021
31 Oct 10, 2022
2.5k Dec 26, 2022
24 May 16, 2022
12 Jul 08, 2022
800 Dec 31, 2022
8 Oct 15, 2021
0 Oct 06, 2021