Domain randomization for neural network classification
Abstract Large data requirements are often the main hurdle in training neural networks. Convolutional neural network (CNN) classifiers in particular require tens of thousands of pre-labeled images per category to approach human-level accuracy, while often failing to generalized to out-of-domain test...
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
SpringerOpen
2021-07-01
|
| Series: | Journal of Big Data |
| Subjects: | |
| Online Access: | https://doi.org/10.1186/s40537-021-00455-5 |
| _version_ | 1819145551600418816 |
|---|---|
| author | Svetozar Zarko Valtchev Jianhong Wu |
| author_facet | Svetozar Zarko Valtchev Jianhong Wu |
| author_sort | Svetozar Zarko Valtchev |
| collection | DOAJ |
| description | Abstract Large data requirements are often the main hurdle in training neural networks. Convolutional neural network (CNN) classifiers in particular require tens of thousands of pre-labeled images per category to approach human-level accuracy, while often failing to generalized to out-of-domain test sets. The acquisition and labelling of such datasets is often an expensive, time consuming and tedious task in practice. Synthetic data provides a cheap and efficient solution to assemble such large datasets. Using domain randomization (DR), we show that a sufficiently well generated synthetic image dataset can be used to train a neural network classifier that rivals state-of-the-art models trained on real datasets, achieving accuracy levels as high as 88% on a baseline cats vs dogs classification task. We show that the most important domain randomization parameter is a large variety of subjects, while secondary parameters such as lighting and textures are found to be less significant to the model accuracy. Our results also provide evidence to suggest that models trained on domain randomized images transfer to new domains better than those trained on real photos. Model performance appears to remain stable as the number of categories increases. |
| first_indexed | 2024-12-22T12:59:50Z |
| format | Article |
| id | doaj.art-ff752bb0d9584aa2b2052aef2bd5fafa |
| institution | Directory Open Access Journal |
| issn | 2196-1115 |
| language | English |
| last_indexed | 2024-12-22T12:59:50Z |
| publishDate | 2021-07-01 |
| publisher | SpringerOpen |
| record_format | Article |
| series | Journal of Big Data |
| spelling | doaj.art-ff752bb0d9584aa2b2052aef2bd5fafa2022-12-21T18:25:01ZengSpringerOpenJournal of Big Data2196-11152021-07-018111210.1186/s40537-021-00455-5Domain randomization for neural network classificationSvetozar Zarko Valtchev0Jianhong Wu1Laboratory of Industrial and Applied Mathematics, York UniversityLaboratory of Industrial and Applied Mathematics, York UniversityAbstract Large data requirements are often the main hurdle in training neural networks. Convolutional neural network (CNN) classifiers in particular require tens of thousands of pre-labeled images per category to approach human-level accuracy, while often failing to generalized to out-of-domain test sets. The acquisition and labelling of such datasets is often an expensive, time consuming and tedious task in practice. Synthetic data provides a cheap and efficient solution to assemble such large datasets. Using domain randomization (DR), we show that a sufficiently well generated synthetic image dataset can be used to train a neural network classifier that rivals state-of-the-art models trained on real datasets, achieving accuracy levels as high as 88% on a baseline cats vs dogs classification task. We show that the most important domain randomization parameter is a large variety of subjects, while secondary parameters such as lighting and textures are found to be less significant to the model accuracy. Our results also provide evidence to suggest that models trained on domain randomized images transfer to new domains better than those trained on real photos. Model performance appears to remain stable as the number of categories increases.https://doi.org/10.1186/s40537-021-00455-5Domain randomizationSynthetic image generationNeural network classifiers |
| spellingShingle | Svetozar Zarko Valtchev Jianhong Wu Domain randomization for neural network classification Journal of Big Data Domain randomization Synthetic image generation Neural network classifiers |
| title | Domain randomization for neural network classification |
| title_full | Domain randomization for neural network classification |
| title_fullStr | Domain randomization for neural network classification |
| title_full_unstemmed | Domain randomization for neural network classification |
| title_short | Domain randomization for neural network classification |
| title_sort | domain randomization for neural network classification |
| topic | Domain randomization Synthetic image generation Neural network classifiers |
| url | https://doi.org/10.1186/s40537-021-00455-5 |
| work_keys_str_mv | AT svetozarzarkovaltchev domainrandomizationforneuralnetworkclassification AT jianhongwu domainrandomizationforneuralnetworkclassification |