QuasiVSD: efficient dual-frame smoke detection

Smoke is a typical symptom of early fire, and the appearance of a large amount of abnormal smoke usually indicates an impending abnormal accident. A smart smoke detection method can substantially reduce damage caused by fires in cities, factories and forests, it is also an important component of intelligent surveillance system. However, existing image-based detection methods often suffer from the lack of dynamic information, and video-based methods are usually computing-expensive because more input images need to be processed. In this work, we propose a novel and efficient Quasi Video Smoke Detector (QuasiVSD) to bridge the gap between image-based and video-based smoke detection. By regarding an unannotated image as reference, QuasiVSD can obtain motion-aware attention from just two frames. Moreover, Weakly Guided Attention Module is designed to further refine the feature representation for smoke regions. Finally, extensive experiments on real-world dataset show that our QuasiVSD achieves clear improvements against the image-based best competitors (CenterNet) by 4.71 with almost same parameters and FLOPs. And the computational complexity of QuasiVSD is just a fraction of that of general video understanding framework. Code will be available at: https://github.com/Caoyichao/VSDT.