Dense and Occlusion-Robust Multi-view Stereo for Unstructured Videos
Dense and Occlusion-Robust Multi-view Stereo for Unstructured Videos
复制标题
DOI：
10.1109/crv.2016.42
复制
发表时间：
2016-06
期刊：
影响因子：
--
通讯作者：
Jian Wei;Benjamin Resch;H. Lensch
中科院分区：
文献类型：
--
作者：
Jian Wei;Benjamin Resch;H. Lensch
我们提出了一种高效的多视点立体系统，用于从任意摄像机轨迹的视频中密集重建均匀区域。大多数技术在恢复无纹理区域方面存在困难，遮挡的鲁棒性较弱，并且在密集采样帧提供的大量冗余数据上效率较低。我们的关键思想是使用物体边缘的可测量深度来恢复封闭的几何形状，假设表面光滑，这适用于大多数场景。边缘处的深度值是通过考虑单独的观察光线来计算的，以允许明显的不连续。然后我们使用透视扩散来创建边缘之间的光滑表面。通过检测大的深度标准偏差，利用其他视图的边缘深度近似，有效地消除了前背景边缘之间的错误插值。为了实现可靠的失效，采用先进的评分评估和双尺度图像采样来提高边缘深度的准确性。最后，我们通过在视图中传播最近的有效深度来填充洞并纠正错误的凹陷表面。像素级操作支持高并行性，我们的高质量深度图允许通过点云密集的场景表示。
We present an efficient multi-view stereo system for dense reconstruction of homogeneous areas from videos with arbitrary camera trajectories. Most techniques have difficulties in recovering textureless areas, weak robustness for occlusion, and low efficiency on the massive redundant data provided by densely sampled frames. Our key idea is to use the measurable depth at object edges to recover the enclosed geometries assuming smooth surfaces, which is appropriate for most scenes. The depth values at edges are calculated by considering individual viewing rays to allow for sharp discontinuities. Then we employ perspective diffusion to create smooth surfaces between edges. The wrong interpolants between fore-and background edges are effectively invalidated by detecting large depth standard deviation which is approximated using the edge depth from other views. For reliable invalidation, the accuracy of edge depth is improved by advanced score evaluation and two-scale image sampling. Finally, we fill up holes and correct the wrongly sunken surfaces by propagating the closest valid depth across views. The pixel-level operations throughout support high parallelism, and our high-quality depth maps allow dense scene representation by point clouds.