Synethic dataset for differential OGI images and wind tunnel dataset

Fluid Flow Dataset

1. Introduction

A total of three datasets are included here, namely, (1) the synthetic dataset used for model training as mentioned in the paper, (2) the wind tunnel dataset used to validate the segmentation performance of the model, including the corresponding manually labeled labels, and (3) the original wind tunnel experiment dataset.

2. Synthetic differential image dataset for segmentation

To address the lack of semantic segmentation datasets for infrared fluid flow imagery, we created a synthetic dataset from the ScalarFlow dataset, focusing on pixel-level labels suitable for neural network training. Our process, illustrated as shown in the paper, includes generating realistic noise by capturing images under controlled conditions with an Optical Gas Imaging (OGI) camera, followed by image subtraction, normalization, and merging with ScalarFlow data. This method ensures the inclusion of real-world disturbances such as camera jitter effects, enhancing the dataset's robustness and applicability. The dataset, enriched with various data augmentation techniques, comprises over 30,000 images split into training, validation, and testing sets, catering to the rigorous demands of practical applications in fluid dynamics analysis.

3. Wind tunnel dataset

To enable the determination of velocities of fluid flow by using optical flow algorithms, a wind tunnel data set that includes fluid images captured by different cameras was recorded. 

The fluid flow is created in the wind tunnel and generated by different substances, i.e., dry ice, smoke matches, or paraffin oil. In addition, velocity data collected from the 3D ultrasonic anemometer were used as a reference to evaluate the performance and accuracy of the optical flow algorithms. The fluid flow rate was set at three different velocities in Euclidean space, i.e., 0.7 m/s,  1.4 m/s, and 2.0 m/s.

This dataset is captured by using a wind tunnel, the OGI camera FLIR GF320 and a 3D anemometer for obtaining reference flow velocities. Below is the information of the used camera.

FLIR GF320 Camera Info

Parameter

Value

Spectral Range

3.2 – 3.4 μm

Standard Temperature Range

–20°C to +350°C

Accuracy

 ±1 °C for 0 °C to 100 °C; ±2% > 100 °C

Lenses

24° × 18°

Resolution

320 × 240 Pixel

The dataset consists of three parts, i.e., OGI images of fluids/smoke generated from three different substances.

1. Smoke matches dataset 

Id

Substances

Velocity in m/s

Group

Number of frames

1
Smoke matches
0.7
1
143

2
Smoke matches
0.7
2
597

3
Smoke matches
1.4
1
145

4
Smoke matches
1.4
2
598

5
Smoke matches
2.0
1
145

6
Smoke matches
2.0
2
596

2. Paraffin oil dataset

Id

Substances

Velocity in m/s

Group

Number of frames

7

Paraffin oil

0.7
1
597

8
Paraffin oil
0.7
2
597

9
Paraffin oil
1.4
1
597

10
Paraffin oil
1.4
2
597

11
Paraffin oil
2.0
1
597

12
Paraffin oil
2.0
2
597

3. Dry ice dataset

Id

Substances

Velocity in m/s

Group

Number of frames

13
Dry ice
0.7
1
598

14
Dry ice
0.7
2
597

15
Dry ice
1.4
1
595

16
Dry ice
1.4
2
596

17
Dry ice
2.0
1
185

18
Dry ice
2.0
2
628

We also provide the corresponding 3D anemometer data, which allows the user to convert the pixel displacement from the image to the actual flow rate, as shown in below.

Velocities in m/s and pixel

Velocity in m/s

Settings of WindChannel

 Velocity in pixel

0.7

1.88
4.57

1.4
2.50
9.14

2.0
3.06
13.06

4. Wind tunnel segmentation dataset

This part of the dataset is from the dry ice dataset portion of the wind tunnel test dataset described above. And labels are generated by manual labeling for evaluating the performance of the image segmentation model in real-world scenarios, a total of 100 differential images and 100 labels.