Robust Cross-Scene Foreground Segmentation in Surveillance Video

Training only one deep model for large-scale cross-scenevideo foreground segmentation is challenging due to the off-the-shelf deep learning based segmentor relies on scene-specific structural information. This results in deep mod-els that are scene-biased and evaluations that are scene-influenced.In this paper, we integrate dual modalities(foregrounds’ motion and appearance), and then eliminat-ing features without representativeness of foreground throughattention-module-guided selective-connection structures. It isin an end-to-end training manner and to achieve scene adap-tation in the plug and play style. Experiments indicate theproposed method significantly outperforms the state-of-the-art deep models and background subtraction methods in un-trained scenes – LIMU and LASIESTA. (Codes and datasetwill be available after the anonymous stage.)