Abstract
This paper considers image change detection with only a small number ofsamples, which is a significant problem in terms of a few annotationsavailable. A major impediment of image change detection task is the lack oflarge annotated datasets covering a wide variety of scenes. Change detectionmodels trained on insufficient datasets have shown poor generalizationcapability. To address the poor generalization issue, we propose using simpleimage processing methods for generating synthetic but informative datasets, anddesign an early fusion network based on object detection which could outperformthe siamese neural network. Our key insight is that the synthetic data enablesthe trained model to have good generalization ability for various scenarios. Wecompare the model trained on the synthetic data with that on the real-worlddata captured from a challenging dataset, CDNet, using six different test sets.The results demonstrate that the synthetic data is informative enough toachieve higher generalization ability than the insufficient real-world data.Besides, the experiment shows that utilizing a few (often tens of) samples tofine-tune the model trained on the synthetic data will achieve excellentresults.