基于非参数化采样的单幅图像深度估计-计算机应用研究.PDFVIP

优先出版 计 算 机 应 用 研 究 第33 卷 基于非参数化采样的单幅图像深度估计* 朱 尧，喻 秋 (武汉大学 计算机学院，武汉 430072) 摘 要：针对传统单幅图像深度估计线索不足及深度估计精度不准的问题，提出一种基于非参数化采样的单幅图像深 度估计方法。该方法利用非参数化的学习手段，将现有 RGBD 数据集中的深度信息迁移到输入图像中去。首先计算输 入图像和现有 RGBD 数据集多尺度的高层次图像特征；然后，在现有 RGBD 数据集中，基于高层次的图像特征通过 kNN 最近邻搜索找到若干与输入图像特征最匹配的候选图像，并将这些候选图像通过 SIFT 流形变到输入图像进行对 齐。最后，对候选深度图进行插值和平滑等优化操作便可以得到最后的深度图。实验结果表明，与现有算法相比，该 方法估计得到的深度图精度更高，对输入图像的整体结构保持得更好。 关键词：单幅图像；深度估计；非参数化采样；RGBD 数据集 中图分类号：TP391.41 Non-parametric sampling based depth estimation from single image Zhu Yao, Yu Qiu (School of Computer, Wuhan University, Wuhan 430072, China) Abstract: Given the insufficient depth cues as well as the inaccurate depth estimation of single images, this paper presented a non-parametric sampling-based depth estimation method . This method transferred the depth information of the images in current RGBD database to the input single image by using the non-parametric sampling way. First, the algorithm computed the high- level multiscale image features of the input and RGBD images. Then, in the existing RGBD dataset, it found a number of candidate images and their respective depth maps which were most matched to the input image in feature space by kNN searching based on the high-level image features. After that, it warped the candidate images and depth maps to align them to the input image via SIFT flow. Finally, it optimized the candidate depth images by applying interpolating and smoothing operations and the final depth map can be achi