Unsupervised learning of Particle Image Velocimetry

This repository contains materials for ISC workshop paper Unsupervised learning of Particle Image Velocimetry.

Introduction

Particle Image Velocimetry (PIV) is a classical flow estimation problem which is widely considered and utilised, especially as a diagnostic tool in experimental fluid dynamics and the remote sensing of environmental flows. We present here what we believe to be the first work which takes an unsupervised learning based approach to tackle PIV problems. The proposed approach is inspired by classic optical flow methods. Instead of using ground truth data, we make use of photometric loss between two consecutive image frames, consistency loss in bidirectional flow estimates and spatial smoothness loss to construct the total unsupervised loss function. The approach shows significant potential and advantages for fluid flow estimation. Results presented here demonstrate that is outputs competitive results compared with classical PIV methods as well as supervised learning based methods for a broad PIV dataset, and even outperforms these existing approaches in some difficult flow cases.

Sample results

Syethetic data: samples from PIV dataset

• Backstep flow

• Surface Quasi Geostrophic (SQG) flow

Real experimental data: particle Images from PIV challenge

• Jet Flow

From left to right: Particle images, UnLiteFlowNet-PIV(trained by full integrated loss) output, PIV-LiteNetFlow output

Unsupervised Loss

Dataset

The dataset used in this work is obtained from the work below:

• PIV dataset (9GB)

Shengze Cai, Shichao Zhou, Chao Xu, Qi Gao. 
Dense motion estimation of particle images via a convolutional neural network, Exp Fluids, 2019

• JHTDB

Y. Li, E. Perlman, M. Wan, Y. Yang, R. Burns, C. Meneveau, R. Burns, S. Chen, A. Szalay & G. Eyink. 
A public turbulence database cluster and applications to study Lagrangian evolution of velocity increments in turbulence. Journal of Turbulence 9, No. 31, 2008.

Prerequisite

• cuda (v10.1)

• pytorch (v1.5.0)

• sklearn (v0.22.2)

• livelossplot

  pip install livelossplot

• flowiz

  Library to visualize .flo files

  pip install flowiz -U

• GPUtil

  pip install GPUtil

Training

To train from scratch:

1. Download the PIV dataset, remove the current data in the folder sample_data and extract new data into it.

2. Run the scripts with --train argument:

   python main.py --train

3. Trained model will be saved in the same folder. (A checkpoint is generated every 5 epochs in default during training)

Trained model

The trained model UnsupervisedLiteFlowNet_pretrained.pt is available in the folder models.

Testing

The data samples for test use are in the folder sample_data.

Test and visualize the sample data results with the pretrained model using:

python main.py --test

Citation