fnctId=thesis,fnctNo=374
Automatic Generation Method for Urban Change Detection Dataset
- Link
- https://doi.org/10.1109/ACCESS.2026.3659083
- Writer
- 고대철
- Author
- JINSU HA, BYEONG HAK KIM, DAE-CHEOL KO, HONG-IN WON, JINSEOK JANG
- Publication matter
- IEEE Access
- Publication Date
- 2026-01
- Korean Abstract
- English Abstract
- This study investigates an inpainting-based data generation approach for urban change
detection and analyzes its impact on the model generalization performance. The proposed method
constructs temporally consistent before and after image pairs by applying inpainting to existing images,
thereby enabling seamless integration with existing change detection frameworks. Rather than replacing
real-world datasets, the generated data were designed to increase the training diversity and mitigate datasetspecific
biases. Comprehensive experiments were conducted using multiple representative change detection
architectures across different datasets. The results show that incorporating the generated data alongside
real training samples consistently improves cross-dataset generalization performance. In particular, models
trained on S2Looking achieved up to a 14.0 percentage-point improvement in F1-score on the unseen
WHU-CD dataset, depending on the model architecture. By contrast, training with the generated data alone
leads to limited generalization when evaluated on real test sets. These findings indicate that inpainting-based
synthetic data can serve as a complementary training resource for enhancing the robustness of urban change
detection. Future work will focus on reducing the domain gap between synthetic and real data by extending
the proposed framework beyond simple object removal to incorporate more diverse change scenarios that
consider background context and temporal variations, such as seasonal appearance changes.
- Text
- File