Siamese partial change network for accurate and fast image change detection

Current state-of-the-art change detection networks depend on siamese CNN backbones to extract powerful feature representations in multi dimensions. However, architectures of existing siamese network methods are bloated and timeconsuming because they simply use whole backbones mentioned above in siamese architecture and roughly fuse all the information. The important step of the task is to design an effective feature extractor and fusion module that can enhance the performance of the corresponding change maps. In this paper, we propose a novel framework named Siamese Partial Change Network (SPCN) for fast and accurate image change segmentation. On the one hand, we modify backbones to extract change features partially in small resolution which are lightweight and more suitable for change detection tasks. On the other hand, we explore an alternative to integrate and refine change features for better results. Extensive experiments on CDnet 2014 dataset show that our proposed model not only outperforms previous state-of-the-art change detection methods, but also runs much faster than most existing models.