Methodology for selecting hyperparameters of a neural network model in optical navigation tasks


Аuthors

Doroshev A. S.1*, Shelomanov D. A.2**

1. Moscow Experimental Design Bureau “Mars”, 1-st Shemilovsky lane 16, building 2, Moscow, 127473, Russia
2. Moscow Aviation Institute (National Research University), 4, Volokolamskoe shosse, Moscow, А-80, GSP-3, 125993, Russia

*e-mail: doroshevas@mai.ru
**e-mail: shelomanovda@mai.ru

Abstract

This article presents a technique for optimizing hyperparameters of a neural network model in tasks of autonomous optical navigation of unmanned aerial vehicles (UAVs) using Simultaneous Localization and Mapping (SLAM) methods. The problem of estimating the distance to objects based on monocular camera data is considered. To solve the problem, a hyperparameter selection technique is proposed, including a consistent narrowing of the hyperparameter search area and the use of Bayesian optimization.
The main stages of the methodology:
1) Defining the initial search area of hyperparameters;
2) Formation of the objective function and the direction of optimization.
3) Conducting a series of experiments;
4) Narrowing the search area for hyperparameters based on the analysis of the coefficient of variation of parameters;
5) Repeat training until the optimal result is achieved.
The application of the proposed methodology made it possible to increase the accuracy of model training while maintaining a fixed amount of trials for training experiments. To test the effectiveness, the architectures of encoder-decoder models with a pre-trained Mobilenet model and NYU Depth Dataset V2 data were used. The training was carried out using the target loss function, which includes metrics on pixel difference, structural similarity and image gradients.
The results showed that the proposed technique improved the prediction of the depth of monocular images, and was also tested on tasks with other architectures (MNIST, ResNet18), confirming the universality of the approach. For example, for ResNet18, the accuracy on the test sample increased by 1.08%.
Optimization of hyperparameters has demonstrated effectiveness in improving the accuracy of models while maintaining a fixed trials amount, which makes the technique especially useful for tasks with limited computing resources.

Keywords:

neural networks, unmanned aerial vehicles, hyperparameters, optical navigation, intelligent systems

References

  1. Forster C., Lynen S., Kneip L., Scaramuzza D. et al. Collaborative monocular slam with multiple micro aerial vehicles. 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2013. P. 3962-3970. DOI: 10.1109/IROS.2013.6696923
  2. Blösch M., Weiss S., Scaramuzza D., Siegwart R. et al. Vision based MAV navigation in unknown and unstructured environments. 2010 IEEE International Conference on Robotics and Automation. IEEE, 2010. P. 21-28. DOI: 10.1109/ROBOT.2010.5509920
  3. Ol'kina D.S. Algorithm of semantic image segmentation for solving the problem of positionong an aircraft on the Earth`s surface. Trudy MAI. 2023. No. 130. (In Russ.). URL: https://trudymai.ru/eng/published.php?ID=174617. DOI: 10.34759/trd-2023-130-18
  4. Buraga A.V., Kostyukov V.M. Comparison of range estimation methods for small unmanned aerial vehicle. Trudy MAI. 2012. No. 53. (In Russ.). URL: https://trudymai.ru/eng/published.php?ID=29624
  5. Li R., Wang S., Gu D. Ongoing evolution of visual SLAM from geometry to deep learning: Challenges and opportunities. Cognitive Computation. 2018. V. 10, No 6. P. 875-889. DOI: 10.1007/s12559-018-9591-8
  6. Wang W., Hu Y., Scherer S. TartanVO: A generalizable learning-based VO. Conference on Robot Learning. 2021. P. 1761-1772. DOI: 10.48550/arXiv.2011.00359
  7. Van Rijn J. N., Hutter F. Hyperparameter importance across datasets. Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 2018. P. 2367-2376. DOI: 10.1145/3219819.3220058
  8. Radosavovic I., Kosaraju R.P., Girshick R., He K., Dollár P. et al. Designing network design spaces. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020. P. 10428-10436. DOI: 10.1109/CVPR42600.2020.01044
  9. Wistuba M., Schilling N., Schmidt-Thieme L. Hyperparameter search space pruning–a new component for sequential model-based hyperparameter optimization. Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part II 15. Springer International Publishing, 2015. P. 104-119. DOI: 10.1007/978-3-319-23525-7_7
  10. Akiba T. et al. Optuna: A next-generation hyperparameter optimization framework. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 2019. P. 2623-2631. DOI: 10.1145/3292500.3330701
  11. Makarov A., Namiot D. Overview of data cleaning methods for machine learning. International Journal of Open Information Technologies. 2023. V. 11, No. 10. P. 70-78. 
  12. Bhat S.F., Birkl R., Wofk D., Wonka P., Müller M. et al. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288. 2023. DOI: 10.48550/arXiv.2302.12288
  13. Alhashim I. High quality monocular depth estimation via transfer learning. arXiv preprint arXiv:1812.11941. 2018. URL: https://doi.org/10.48550/arXiv.1812.11941
  14. Wang Z., Bovik A.C., Sheikh H.R., Simoncelli E.P. et al. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing. 2004. V. 13, No. 4. P. 600-612. DOI: 10.1109/TIP.2003.819861
  15. Kanopoulos N., Vasanthavada N., Baker R.L. Design of an image edge detection filter using the Sobel operator. IEEE Journal of solid-state circuits. 1988. V. 23, No. 2. P. 358-367. URL: https://doi.org/10.1109/4.996
  16. Lee K., Yim J. Hyperparameter optimization with neural network pruning. arXiv preprint arXiv:2205.08695. 2022. DOI: 10.48550/arXiv.2205.08695
  17. Engel J., Koltun V., Cremers D. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence. 2017. V. 40, No. 3. P. 611-625. DOI: 10.1109/TPAMI.2017.2658577
  18. Mokssit S. et al. Deep learning techniques for visual slam: A survey. IEEE Access. 2023. V. 11, P. 20026-20050. DOI: 10.1109/ACCESS.2023.3249661
  19. Bischl B. et al. Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery. 2023. V. 13, No. 2. P. 1484. DOI: 10.1002/widm.1484
  20. Yang L., Shami A. On hyperparameter optimization of machine learning algorithms: Theory and practice. Neurocomputing. 2020. V. 415, P. 295-316. DOI: 10.1016/j.neucom.2020.07.061
  21. Bartz E. et al. Experimental investigation and evaluation of model-based hyperparameter optimization. arXiv preprint arXiv:2107.08761. 2021. DOI: 10.48550/arXiv.2107.08761


Download

mai.ru — informational site MAI

Copyright © 2000-2025 by MAI

 Вход