利用深度传感器大数据的单目图像深度估计

Single View Depth Estimation via RGBD Big Data

2D视频转3D视频是解决3D片源不足的主要方法之一,而单目图像的深度估计是其中的关键步骤.考虑到互联网上不断累积的深度图数据,提出一种基于PatchMatch深度迁移的单目图像深度估计方法.首先利用图像的全局描述符从深度图数据库中检索出近邻图像;然后通过PatchMatch建立输入图像和近邻图像之间像素级稠密对应关系;再根据像素级对应关系将近邻图像的深度图迁移到输入图像上,并采用中值滤波对迁移的深度图进行融合;最后通过三边滤波对融合的深度图进行后处理,进一步提高深度图估计质量并抑制噪声.实验结果表明,与基于尺度不变特征变换流深度迁移方法相比,该方法在改善深度图估计质量的同时提高了计算速度.

2D-to-3D conversion is one way to alleviate the lack of 3D-TV program material. The most important and difficult issue in 2D-to-3D conversion is how to estimate the depth map from a monocular image. This paper proposes a PatchMatch depth transfer method of depth estimation from a monocular image for 2D-to-3D conversion based on RGBD data from internet. First, the proposed method retrieves K-nearest neighbor images from RGBD database using global image descriptors such as GIST features. Then, the proposed method matches the input image to its neighbor images by the PatchMatch method. Third, the proposed method transfers depth maps of neighbor images to the input image and estimates its initial depth by median filtering on these transferred depth maps. Finally, the proposed method refines the initial depth map using tri-lateral filtering, in order to further improve the depth map estimation quality. Experimental results show that the proposed approach can greatly reduce the computation burden with depth map quality improvement compared to the scale invariant feature transform flow based method.