Enhancing Change Detection with Self-Supervised Jigsaw Loss

The ability of humans to adapt is dependent on learning from both labeled data and unlabeled experiences. Since supervised learning is inherently limited and cannot encompass all scenarios, the integration of unsupervised learning is essential since it allows models to uncover underlying patterns and invariances that aid in broader generalization. Inspired by this principle, we present a novel method for tackling change detection that blends supervised semantic segmentation with a complementary self-supervised objective. Specifically, our method not only identifies changed pixels in images, but also enhances spatial reasoning by training the model to solve jigsaw puzzles constructed from shuffled segmentation masks. We introduce Jigsaw Change Detection (JCD), a dual-task architecture designed to leverage spatial structure within the data. The auxiliary deshuffling task encourages the model to learn spatial dependencies, effectively regularizing the primary segmentation process. The integration of these tasks leads to improved performance in change detection. Our extensive evaluations across CDD, SYSU, and NJDS benchmarks demonstrate the effectiveness of our approach, consistently outperforming existing methods. Code is made publicly accessible at https://github.com/berkealgul/JDS.