Bridging the Synthetic-Real Gap: Unsupervised Domain Adaptation for Cross-Domain Image Segmentation

N Abhishek, Akshay Poojary, Varsha Sajjanavar, Apeksha Tangod, Sneha Varur, Channabasappa Muttal

2025

Abstract

It is a challenging problem for cross-domain image segmentation bridging the gap between synthetic and real worlds, which is very relevant given applications in autonomous driving scenarios. This work proposes an effective strategy for solving the problem in unsupervised domain adaptation for cross-domain image segmentation; training the model on the GTA5 dataset and testing it on the Cityscapes. We used the ResNet-101 backbone with DeeplabV3+ and exploited its encoder for feature extraction and an upsampling decoder for effective segmentation. The results show that the approach is quite robust for dealing with domain shifts. Although a domain gap exists between the synthetic and real datasets, it correctly segments complex urban scenes. This work makes segmentation models more accurate and generalizable in real applications by using synthetic training data within an unsupervised learning framework. The two major metrics used to evaluate the work are IoU and mean IoU (mIoU). Our method reached a mIoU of 55.80%, outperforming most state-of-the-art UDA methods for the cross-domain segmentation task.

Paper Citation

in Harvard Style

Abhishek N., Poojary A., Sajjanavar V., Tangod A., Varur S. and Muttal C. (2025). Bridging the Synthetic-Real Gap: Unsupervised Domain Adaptation for Cross-Domain Image Segmentation. In Proceedings of the 3rd International Conference on Futuristic Technology - Volume 3: INCOFT; ISBN 978-989-758-763-4, SciTePress, pages 254-262. DOI: 10.5220/0013613500004664

in Bibtex Style

@conference{incoft25,
author={N Abhishek and Akshay Poojary and Varsha Sajjanavar and Apeksha Tangod and Sneha Varur and Channabasappa Muttal},
title={Bridging the Synthetic-Real Gap: Unsupervised Domain Adaptation for Cross-Domain Image Segmentation},
booktitle={Proceedings of the 3rd International Conference on Futuristic Technology - Volume 3: INCOFT},
year={2025},
pages={254-262},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013613500004664},
isbn={978-989-758-763-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 3rd International Conference on Futuristic Technology - Volume 3: INCOFT
TI - Bridging the Synthetic-Real Gap: Unsupervised Domain Adaptation for Cross-Domain Image Segmentation
SN - 978-989-758-763-4
AU - Abhishek N.
AU - Poojary A.
AU - Sajjanavar V.
AU - Tangod A.
AU - Varur S.
AU - Muttal C.
PY - 2025
SP - 254
EP - 262
DO - 10.5220/0013613500004664
PB - SciTePress