Video Saliency Object Detection with Motion Quality Compensation

Video saliency object detection is one of the classic research problems in computer vision, yet existing works rarely focus on impact input quality model performance. As optical flow a key for video models, its significantly affects Traditional models only calculate between two consecutive frames, ignoring motion state objects over period time, leading to low-quality and reduced performance models. Therefore, this paper proposes new that improves by expanding perception range uses high-quality enhance Experimental results datasets show proposed can improve quality, with S-M values DAVSOD dataset increasing about 39%, 49%, 44% compared such as PWCNet, SpyNet, LFNet. In addition, experiments fine-tuning benchmark LIMS demonstrate improving further