Ma Zhuang, Cen Zhaofeng, Li Xiaotong
Appl Opt. 2017 Aug 10;56(23):6603-6610. doi: 10.1364/AO.56.006603.
In this paper, we propose an optimized algorithm to estimate the depth information in the 4D light field data. Our scheme has the advantage of conciseness compared to the traditional epipolar-plane image analysis method. First, we have analyzed the depth resolution properties of light field data not mentioned by the previous researchers. In the depth estimation process, epipolar analysis is confined in a small range to reduce the running time, combining with a regression test to reduce estimation error. Occlusion condition is especially dealt with by recognizing object margin. To test the accuracy of our algorithm, we use a benchmark dataset to evaluate the output depth result. We get a competitive result in the estimation error evaluation and prevailing runtime result compared to that of baseline algorithms. Owing to the high performance, this algorithm can be used in real-time depth recognition with the aid of parallel computing.
在本文中，我们提出了一种优化算法来估计四维光场数据中的深度信息。与传统的极平面图像分析方法相比，我们的方案具有简洁的优点。首先，我们分析了先前研究人员未提及的光场数据的深度分辨率特性。在深度估计过程中，极线分析被限制在一个小范围内以减少运行时间，并结合回归测试以减少估计误差。通过识别物体边缘来特别处理遮挡情况。为了测试我们算法的准确性，我们使用一个基准数据集来评估输出的深度结果。与基线算法相比，我们在估计误差评估和普遍的运行时间结果方面都取得了有竞争力的结果。由于高性能，该算法可借助并行计算用于实时深度识别。
