Faster R-CNN model learning on synthetic images
Błażej Łach
blazej.lach@pollub.edu.plPolitechnika Lubelska (Poland)
Edyta Łukasik
Lublin University of Technology (Poland)
Abstract
Machine learning requires a human description of the data. The manual dataset description is very time consuming. In this article was examined how the model learns from artificially created images, with the least human participation in describing the data. It was checked how the model learned on artificially produced images with augmentations and progressive image size. The model has achieve up to 3.35 higher mean average precision on syntetic dataset in the training with increasing images resolution. Augmentations improved the quality of detection on real photos. The production of artificially generated training data has a great impact on the acceleration of prepare training, because it does not require as much human resources as normal learning process.
Keywords:
computer vision, synthetic images, Faster R-CNN, deep learningReferences
S. Ren et al.: Faster R-CNN: Towards Real-Time Ob-ject Detection with Region Proposal Networks, IEEETransactions on Pattern Analysis and Machine Intel-ligence, t. 39, nr 6, s. 1137–1149, cze. 2017, https://doi.org/10.1109/TPAMI.2016.2577031.
DOI: https://doi.org/10.1109/TPAMI.2016.2577031
Google Scholar
R. Girshick: Fast R-CNN, [W:] 2015 IEEE InternationalConference on Computer Vision (ICCV), 2015, https://doi.org/10.1109/ICCV.2015.169.
DOI: https://doi.org/10.1109/ICCV.2015.169
Google Scholar
H. Jiang, E. Learned-Miller: Face Detection with theFaster R-CNN, [W:] 2017 12th IEEE International Con-ference on Automatic Face & Gesture Recognition (FG2017), 2017, https://doi.org/10.1109/FG.2017.82.
DOI: https://doi.org/10.1109/FG.2017.82
Google Scholar
B. Zoph et al.: Learning Data Augmentation Strate-gies for Object Detection, 2019, https://arxiv.org/pdf/1906.11172.pdf.
Google Scholar
E. Cubuk et al.: AutoAugment: Learning AugmentationStrategies from Data, [W:] 2019 IEEE/CVF Conferenceon Computer Vision and Pattern Recognition (CVPR),2019, https://doi.org/10.1109/CVPR.2019.00020.
DOI: https://doi.org/10.1109/CVPR.2019.00020
Google Scholar
H. Clever et al.: Bodies at Rest: 3D Human Pose andShape Estimation from a Pressure Image using Synthet-ic Data, [W:] 2020 IEEE/CVF Conference on Com-puter Vision and Pattern Recognition (CVPR), 2020,https://doi.org/10.1109/CVPR42600.2020.00625.
DOI: https://doi.org/10.1109/CVPR42600.2020.00625
Google Scholar
M. Danielczuk et al.: Segmenting Unknown 3D Objectsfrom Real Depth Images using Mask R-CNN Trained onSynthetic Data, [W:] 2019 International Conference onRobotics and Automation (ICRA), 2019, https://doi.org/10.1109/ICRA.2019.8793744.
DOI: https://doi.org/10.1109/ICRA.2019.8793744
Google Scholar
A. Pashevich et al.:Learning to Augment Syn-thetic Images for Sim2Real Policy Transfer, [W:]2019 IEEE/RSJ International Conference on IntelligentRobots and Systems (IROS), 2019, https://doi.org/10.1109/IROS40897.2019.8967622.
DOI: https://doi.org/10.1109/IROS40897.2019.8967622
Google Scholar
M. Abadi et al.:TensorFlow: Large-Scale MachineLearning on Heterogeneous Distributed Systems, 2016, https://arxiv.org/pdf/1603.04467.pdf.
Google Scholar
A. Buslaev et al.: Albumentations: fast and flexibleimage augmentations, Information, t. 11, nr 2, s. 125, luty 2020, https://doi.org/10.3390/info11020125.4
Google Scholar
Authors
Edyta ŁukasikLublin University of Technology Poland
Statistics
Abstract views: 293PDF downloads: 232
License
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
