Spatial-temporal Fusion Network for Fast Video Shadow Detection

Existing video shadow detectors often need postprocessing or additional input to perform better, thereby degrading their video shadow detection speed. In this work, we present a novel spatial-temporal fusion network (STF-Net), which can efficiently detect shadows in videos with real-time speed (30FPS) and postprocessing-free. Our STF-Net is based solely on an attention-based spatial-temporal fusion block, equipping with recurrence and CNNs entirely. Experimental results on ViSha validation dataset show that our network exceeds state-of-the-art methods quantitatively and qualitatively.