The combination of synthetic datasets with real-world datasets has gained importance in improving the performance of computer vision models, especially in cases where direct access to real-world data is challenging or impossible to annotate. This thesis work explores the application of domain adaptation techniques in order to fix the domain shift problem between synthetic and real-world datasets when training convolutional neural networks. Using real datasets, which are UAVID and ACDC, and synthetic datasets, which are SELMA and Syndrone, this work shows how such domain adaptation approaches enable a synthetic data-trained model to generalize well under real-life settings. Moreover, the shift caused by different view angles is explored by using images collected from drone and car views. The research work consists of two stages: standard supervised training and domain adaptation. Initial experiments reveal significant performance degradation when models trained on synthetic datasets are directly tested on real-world data due to inherent distributional discrepancies. To solve such cases of differences, several domain adaptation approaches are used; they report good model accuracy improvements. The results show the effectiveness of domain adaptation in closing the gap between different types of data, reaffirming thereby the importance of noise handling, data cleaning, and preprocessing strategies in very general and large datasets. It thus highlights the applicability of domain adaptation techniques to the potential creation of strong models that could withstand real-world challenges in areas like autonomous navigation, healthcare, and environmental monitoring.
The combination of synthetic datasets with real-world datasets has gained importance in improving the performance of computer vision models, especially in cases where direct access to real-world data is challenging or impossible to annotate. This thesis work explores the application of domain adaptation techniques in order to fix the domain shift problem between synthetic and real-world datasets when training convolutional neural networks. Using real datasets, which are UAVID and ACDC, and synthetic datasets, which are SELMA and Syndrone, this work shows how such domain adaptation approaches enable a synthetic data-trained model to generalize well under real-life settings. Moreover, the shift caused by different view angles is explored by using images collected from drone and car views. The research work consists of two stages: standard supervised training and domain adaptation. Initial experiments reveal significant performance degradation when models trained on synthetic datasets are directly tested on real-world data due to inherent distributional discrepancies. To solve such cases of differences, several domain adaptation approaches are used; they report good model accuracy improvements. The results show the effectiveness of domain adaptation in closing the gap between different types of data, reaffirming thereby the importance of noise handling, data cleaning, and preprocessing strategies in very general and large datasets. It thus highlights the applicability of domain adaptation techniques to the potential creation of strong models that could withstand real-world challenges in areas like autonomous navigation, healthcare, and environmental monitoring.
SYNTHETIC TO REAL DOMAIN ADAPTATION FOR WEATHER CLASSIFICATION
KOCA, GÖZDENUR
2024/2025
Abstract
The combination of synthetic datasets with real-world datasets has gained importance in improving the performance of computer vision models, especially in cases where direct access to real-world data is challenging or impossible to annotate. This thesis work explores the application of domain adaptation techniques in order to fix the domain shift problem between synthetic and real-world datasets when training convolutional neural networks. Using real datasets, which are UAVID and ACDC, and synthetic datasets, which are SELMA and Syndrone, this work shows how such domain adaptation approaches enable a synthetic data-trained model to generalize well under real-life settings. Moreover, the shift caused by different view angles is explored by using images collected from drone and car views. The research work consists of two stages: standard supervised training and domain adaptation. Initial experiments reveal significant performance degradation when models trained on synthetic datasets are directly tested on real-world data due to inherent distributional discrepancies. To solve such cases of differences, several domain adaptation approaches are used; they report good model accuracy improvements. The results show the effectiveness of domain adaptation in closing the gap between different types of data, reaffirming thereby the importance of noise handling, data cleaning, and preprocessing strategies in very general and large datasets. It thus highlights the applicability of domain adaptation techniques to the potential creation of strong models that could withstand real-world challenges in areas like autonomous navigation, healthcare, and environmental monitoring.File | Dimensione | Formato | |
---|---|---|---|
Koca_Gozdenur.pdf
accesso aperto
Dimensione
3.91 MB
Formato
Adobe PDF
|
3.91 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/83214