基于时空显著性检测的3D视频生成方法

为了解决3D视频生成方法在获取深度线索方面的难题,使用显著图代替深度图进行3D视频生成。显著图和深度图在性质上有所不同,但是显著图是通过视觉注意力分析得到的,因此也可以给予人眼良好的感官体验。为了得到更适合进行视频内容转换的显著图,将时间信息融入到了深度学习模型当中。通过实验证明了本文方法在两个广泛使用的视频显著性数据集上拥有很好的表现力,所生成的3D内容也具有良好的视觉效果,证明了基于显著性检测的3D视频生成方法具有一定的可行性。

3D Video Generation Method Base on Spatiotemporal Saliency Detection

3D video generation method remains an ongoing challenge due to the difficulty in extracting depth cue. To address this issue, a method for 3D video generation which replaces the depth map with the saliency map was proposed. Although the saliency map and depth map are different in characteristic, the saliency map is obtained by visual attention analysis, therefore an immersive 3D effort of human eye could be given through saliency map. In order to predict adapted saliency map for 3D video conversion, the temporal information for detection model was employed by this proposed method. Experimental results demonstrated that the proposed model gains satisfactory performance over two public video saliency datasets, and the generated 3D content has the immersive visual effects.