# Deep Learning Optical Flow with Compound Loss for Dense Fluid Motion Estimation

^{*}

## Abstract

**:**

## 1. Introduction

## 2. Approach and Materials

#### 2.1. Flownet2 Network Structure

#### 2.2. Compound Loss Function

#### 2.2.1. RMSE Loss Function

#### 2.2.2. AAE Loss Function

#### 2.2.3. Div-Curl Smooth Loss Function

#### 2.2.4. Compound Form

#### 2.3. Synthetic PIV Data Set

## 3. Results and Analysis

#### 3.1. Index of Performance Measurement

#### 3.2. Cosine Similarity Verification of Compound Loss Function

^{9}parameters, this gradient is usually a hyperdimensional vector. Similar to Formula (2) for calculating the cosine value of the angle between two vectors, Formula (9) can be used to measure the cosine similarity of two hyperdimensional vectors.

_{RMSE}and L

_{AAE}, while the cosine similarity between L

_{S}and L

_{RMSE}or L

_{AAE}is low but still over 0, which indicates that there is no confrontation between the tasks. This shows that all three parts of the compound loss function play a positive role in the optimization of the target task.

#### 3.3. Sensitivity Analysis of Compound Loss Function Parameters

^{9}model parameters, which adds difficulty to the training of the model. Therefore, in this part, we reduce the synthetic data set applied in Section 2.3 to 1/10, and verify the effect of parameter selection with a data set of about 1000 pairs of images. At the same time, because there are too few samples in the training data set, the final effect of the training is worse than the final result in Section 3.4. The purpose of this experiment is to compare the impact of different parameter choices on the model training effect. The experimental results are shown in Table 3 below.

_{RMSE}and adjusting each part to a similar order of magnitude. However, hyperparameter tuning of compound loss functions is still an issue that is difficult to fully discuss.

#### 3.4. Results of Compound Form of Loss Function on Training Data Set

_{RMSE}function and the various parts of the composite loss function on the large training effect. The parameter settings of the composite loss function are given in Section 2.2 and verified in Section 3.3. The experimental results show that incorporating the angular error into the loss function can not only improve the AAE accuracy of the final calculation result, but also improve the RMSE calculation accuracy to a certain extent. Combining the RMSE loss function with the div-curl loss function alone cannot achieve the desired effect. However, combining the three parts can achieve the best results in the three indicators of RMSE, AAE, and div-curl loss. This shows that curl and divergence, as measures of the local-global information of the flow field, do not contribute to the improvement of the calculation effect when the local information is not accurate enough, but when the single vector angle and amplitude errors are considered, the div-curl loss plays its part.

#### 3.5. Analysis of RMSE Calculation Results

#### 3.6. Analysis of AAE Calculation Results

#### 3.7. Analysis of Div-Curl Error Calculation Results

## 4. Conclusions

## Author Contributions

## Funding

## Data Availability Statement

## Conflicts of Interest

## References

- Raffel, M.; Willert, C.; Wereley, S.; Kompenhans, J. Particle Image Velocimetry: A Practical Guide; Springer: Cham, Switzerland, 2007. [Google Scholar]
- Cai, S. Optical Flow-Based Motion Estimation of Complex Flows; Zhejiang University: Hangzhou, China, 2019. [Google Scholar]
- Wang, J.; Shi, T.; Yu, D.; Teng, D.; Ge, X.; Zhang, Z.; Yang, X.; Wang, H.; Wu, G. Ensemble machine-learning-based framework for estimating total nitrogen concentration in water using drone-borne hyperspectral imagery of emergent plants: A case study in an arid oasis, NW China. Environ. Pollut.
**2020**, 266, 115412. [Google Scholar] [CrossRef] [PubMed] - Shinohara, K.; Sugii, Y.; Aota, A.; Hibara, A.; Tokeshi, M.; Kitamori, T.; Okamoto, K. High-Speed Micro PIV Measurements of Micro Counter-Current Flow. Proc. JSME Annu. Meet.
**2004**, 2004, 111–112. [Google Scholar] [CrossRef] - Westerweel, J. Digital Particle Image Velocimetry: Theory and Application. Ph.D. Thesis, Delft University, Delft, The Netherlands, 1995. [Google Scholar]
- Scarano, F.; Riethmuller, M.L. Iterative Multigrid Approach in PIV Image Processing with Discrete Window Offset. Exp. Fluids
**1999**, 26, 513–523. [Google Scholar] [CrossRef] - Horn, B.; Schunck, B. Determining optical flow. Artif. Intell.
**1981**, 17, 185–203. [Google Scholar] [CrossRef][Green Version] - Corpetti, T.; Memin, E.; Perez, P. Dense Estimation of Fluid Flows. Pattern Anal. Mach. Intell. IEEE Trans.
**2002**, 24, 365–380. [Google Scholar] [CrossRef][Green Version] - Liu, T.S.; Shen, L.X. Fluid flow and optical flow. J. Fluid Mech.
**2008**, 614, 253–291. [Google Scholar] [CrossRef] - Tauro, F.; Tosi, F.; Mattoccia, S.; Toth, E.; Piscopia, R.; Grimaldi, S. Optical Tracking Velocimetry (OTV): Leveraging Optical Flow and Trajectory-Based Filtering for Surface Streamflow Observations. Remote Sens.
**2018**, 10, 2010. [Google Scholar] [CrossRef][Green Version] - Khalid, M.; Pénard, L.; Mémin, E. Optical flow for image-based river velocity estimation. Flow Meas. Instrum.
**2019**, 65, 110–121. [Google Scholar] [CrossRef][Green Version] - Hui, T.W.; Tang, X.; Loy, C.C. LiteFlowNet: A Lightweight Convolutional Neural Network for Optical Flow Estimation. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–23 June 2018. [Google Scholar]
- Fischer, P.; Dosovitskiy, A.; Ilg, E.; Häusser, P.; Hazırbaş, C.; Golkov, V.; Van der Smagt, P.; Cremers, D.; Brox, T. FlowNet: Learning Optical Flow with Convolutional Networks. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 7–13 December 2015. [Google Scholar]
- Sun, D.; Yang, X.; Liu, M.Y.; Kautz, J. PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume. In Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Salt Lake City, UT, USA, 18–23 June 2018. [Google Scholar]
- Teed, Z.; Deng, J. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow; Springer: Berlin/Heidelberg, Germany, 2020. [Google Scholar]
- Ilg, E.; Mayer, N.; Saikia, T.; Keuper, M.; Dosovitskiy, A.; Brox, T. FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017. [Google Scholar]
- Lee, Y.; Yang, H.; Yin, Z. PIV-DCNN: Cascaded deep convolutional neural networks for particle image velocimetry. Exp. Fluids
**2017**, 58, 171. [Google Scholar] [CrossRef] - Kondor, S.; Chan, D.; Sitterle, J. Application of Optical Surface Flow Measurement to Composite Resin Shrinkage. In Proceedings of the ADEA/AADR/CADR Meeting & Exhibition, Orlando, FL, USA, 8–9 March 2006. [Google Scholar]
- Cai, S.; Liang, J.; Gao, Q.; Xu, C.; Wei, R. Particle Image Velocimetry Based on a Deep Learning Motion Estimator. IEEE Trans. Instrum. Meas.
**2020**, 69, 3538–3554. [Google Scholar] [CrossRef] - Cai, S.; Zhou, S.; Xu, C.; Gao, Q. Dense motion estimation of particle images via a convolutional neural network. Exp. Fluids
**2019**, 60, 73. [Google Scholar] [CrossRef] - Dickson, M.C.; Bosman, A.S.; Malan, K.M. Hybridised loss functions for improved neural network generalization. In Proceedings of the Pan-African Artificial Intelligence and Smart Systems: First International Conference, PAAISS 2021, Windhoek, Namibia, 6–8 September 2021; Springer International Publishing: Cham, Switzerland, 2022; pp. 169–181. [Google Scholar]
- Kendall, A.; Gal, Y.; Cipolla, R. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2018, Salt Lake City, UT, USA, 18–22 June 2018; pp. 7482–7491. [Google Scholar]
- Sener, O.; Koltun, V. Multi-task learning as multi-objective optimization. Adv. Neural Inf. Process. Syst.
**2018**, 31, 1–12. [Google Scholar] - Yu, C.D.; Fan, Y.W.; Bi, X.J.; Han, Y.; Kuai, Y.F. Deep particle image velocimetry supervised learning under light conditions. Flow Meas. Instrum.
**2021**, 80, 102000. [Google Scholar] [CrossRef] - Yu, C.; Luo, H.; Fan, Y.; Bi, X.; He, M. A cascaded convolutional neural network for two-phase flow PIV of an object entering water. IEEE Trans. Instrum. Meas.
**2021**, 71, 1–10. [Google Scholar] [CrossRef] - Baker, S.; Scharstein, D.; Lewis, J.P.; Roth, S.; Black, M.J.; Szeliski, R. A Database and Evaluation Methodology for Optical Flow. Int. J. Comput. Vis.
**2011**, 92, 1–31. [Google Scholar] [CrossRef][Green Version] - Yu, T.; Kumar, S.; Gupta, A.; Levine, S.; Hausman, K.; Finn, C. Gradient surgery for multi-task learning. Adv. Neural Inf. Process. Syst.
**2020**, 33, 5824–5836. [Google Scholar]

Method | Train | Test | ||
---|---|---|---|---|

Sintel Clean | Sintel Final | Sintel Clean | Sintel Final | |

FlowNetS | 4.50 | 5.45 | 7.42 | 8.43 |

FlowNetC | 4.31 | 5.87 | 7.28 | 8.81 |

FlowNet2 | 2.02 | 3.54 | 3.96 | 6.02 |

LiteFlowNet | 2.48 | 4.04 | - | - |

PWC-Net | 2.55 | 3.93 | - | - |

Flow Field Type | Max Displacement (Pixel) | Step (Pixel) | Amount |
---|---|---|---|

Linear | ±5~ ± 20 | 0.1 | 4500 |

Rankine | 5~25 | 0.1 | 2400 |

Hamel-Ossen | 5–20 | 0.1 | 1280 |

Rotation | ±5–±20 | 0.1 | 600 |

Membrane | 3 | 0 | 600 |

${\mathit{\lambda}}_{1}$ (L_{RMSE}) | ${\mathit{\lambda}}_{2}$ (L_{AAE}) | ${\mathit{\lambda}}_{3}$ (L_{S}) | RMSE (Pixel) | AAE (°) |
---|---|---|---|---|

1 | 1 | 1 | 4.886 | 17.658 |

1 | 0.1 | 10 | 2.791 | 5.314 |

1 | 0.05 | 10 | 2.368 | 4.254 |

10 | 1 | 10 | 3.081 | 4.346 |

10 | 1 | 1 | 3.427 | 4.891 |

Loss Function | RMSE (Pixel) | AAE (°) | Curl Error | Div Error |
---|---|---|---|---|

L_{RMSE} | 0.201 | 3.73 | 0.084 | 0.065 |

L_{RMSE} + L_{AAE} | 0.193 | 2.56 | 0.046 | 0.038 |

L_{RMSE} + L_{S} | 0.224 | 4.34 | 0.022 | 0.019 |

L_{RMSE} + L_{AAE} + L_{S} | 0.182 | 1.724 | 0.024 | 0.017 |

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. |

© 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Wang, J.; Zhang, Z.; Wang, Z.; Chen, L.
Deep Learning Optical Flow with Compound Loss for Dense Fluid Motion Estimation. *Water* **2023**, *15*, 1365.
https://doi.org/10.3390/w15071365

**AMA Style**

Wang J, Zhang Z, Wang Z, Chen L.
Deep Learning Optical Flow with Compound Loss for Dense Fluid Motion Estimation. *Water*. 2023; 15(7):1365.
https://doi.org/10.3390/w15071365

**Chicago/Turabian Style**

Wang, Jie, Zhen Zhang, Zhijian Wang, and Lin Chen.
2023. "Deep Learning Optical Flow with Compound Loss for Dense Fluid Motion Estimation" *Water* 15, no. 7: 1365.
https://doi.org/10.3390/w15071365