Influence of Synthetic Image Datasets on the Result of Neural Networks for Object Detection
| Autor: | Aleksandr Kniazev, Pavel Slivnitsin, Leonid Mylnikov, Stefan Schlechtweg, Andrey Kokoulin |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | Proceedings of the International Conference on Applied Innovations in IT, Vol 9, Iss 1, Pp 55-60 (2021) |
| Druh dokumentu: | article |
| ISSN: | 2199-8876 82212856 |
| DOI: | 10.25673/36584 |
| Popis: | The goal of the article is research of ways to improve the quality of neural networks object detection. To achieve this goal we suggest to use synthetic image datasets. The algorithm of generating synthetic images, which uses the environment of the detected object, is described in the article. That algorithm could be applied in the control algorithm of the robotic system for luminaire replacement that is based on target object detection. 3D models and 3D camera images of detected objects, backgrounds, noise objects and different effects are used to create realistic images that will increase the quality of predictions. Quality tests were made with synthetic and real datasets. Results show that quality could be increased up to 16%. Ratio of real and synthetic data is 1:4. |
| Databáze: | Directory of Open Access Journals |
| Externí odkaz: |
|