Learning to Detect Fine-Grained Change Under Variant Imaging Conditions

Abstract

Fine-grained change detection under variant imaging conditions is an important and challenging task for high-value scene monitoring in culture heritage. State-of-the-art methods solve this problem by jointly optimizing three related factors, i.e., camera pose difference, illumination variation, and the true minute change of the scene. Their performances are highly dependent on the delicate choice of key parameters, which significantly limits their feasibility in real-world applications. In this paper, we show that after a simple coarse alignment of lighting and camera differences, fine-grained change detection can be reliably solved by a deep network model, which is specifically composed of three functional parts, i.e., camera pose correction network (PCN), fine-grained change detection network (FCDN), and detection confidence boosting. Since our model is properly pre-trained and fine-tuned on both general and specialized data, it exhibits very good generalization capability to produce high-quality minute change detection on real-world scenes under varied imaging conditions. Extensive experiments validate the superior effectiveness and reliability over state-of-the-art methods. We have achieved 67.41% relative F1-measure improvement over the best competitor on real-world benchmark dataset.