针对基于深度学习的多视图立体(Multi-view Stereo,MVS)重建算法内存消耗过大、推理速度慢,以及对病态区域重建效果不佳的问题,提出了一种基于双边网格和融合代价体的轻量级级联的MVS重建网络。首先利用基于双边网格的代价体上采样模块...针对基于深度学习的多视图立体(Multi-view Stereo,MVS)重建算法内存消耗过大、推理速度慢,以及对病态区域重建效果不佳的问题,提出了一种基于双边网格和融合代价体的轻量级级联的MVS重建网络。首先利用基于双边网格的代价体上采样模块将较低分辨率代价体高效地恢复成高分辨率代价体。随着采用轻量级的动态区域卷积和粗粒度代价体融合模块,提升网络对病态区域特征的表示能力以及对场景整体信息和结构信息的感知能力。实验结果表明,该网络在DTU数据集以及Tanks and Temples数据集上均取得了具有竞争性的结果,并且在内存消耗以及推理速度上都显著优于其他方法。展开更多
The mapping method is a forward-modeling method that transforms the irregular surface to horizontal by mapping the rectangular grid as curved; moreover, the wave field calculations move from the physical domain to the...The mapping method is a forward-modeling method that transforms the irregular surface to horizontal by mapping the rectangular grid as curved; moreover, the wave field calculations move from the physical domain to the calculation domain. The mapping method deals with the irregular surface and the low-velocity layer underneath it using a fine grid. For the deeper high-velocity layers, the use of a fine grid causes local oversampling. In addition, when the irregular surface is transformed to horizontal, the flattened interface below the surface is transformed to curved, which produces inaccurate modeling results because of the presence of ladder-like burrs in the simulated seismic wave. Thus, we propose the mapping method based on the dual-variable finite-difference staggered grid. The proposed method uses different size grid spacings in different regions and locally variable time steps to match the size variability of grid spacings. Numerical examples suggest that the proposed method requires less memory storage capacity and improves the computational efficiency compared with forward modeling methods based on the conventional grid.展开更多
文摘针对基于深度学习的多视图立体(Multi-view Stereo,MVS)重建算法内存消耗过大、推理速度慢,以及对病态区域重建效果不佳的问题,提出了一种基于双边网格和融合代价体的轻量级级联的MVS重建网络。首先利用基于双边网格的代价体上采样模块将较低分辨率代价体高效地恢复成高分辨率代价体。随着采用轻量级的动态区域卷积和粗粒度代价体融合模块,提升网络对病态区域特征的表示能力以及对场景整体信息和结构信息的感知能力。实验结果表明,该网络在DTU数据集以及Tanks and Temples数据集上均取得了具有竞争性的结果,并且在内存消耗以及推理速度上都显著优于其他方法。
基金financially supported by the National Natural Science Foundation of China(Nos.41104069 and 41274124)the National 973 Project(Nos.2014CB239006 and 2011CB202402)+1 种基金the Shandong Natural Science Foundation of China(No.ZR2011DQ016)Fundamental Research Funds for Central Universities(No.R1401005A)
文摘The mapping method is a forward-modeling method that transforms the irregular surface to horizontal by mapping the rectangular grid as curved; moreover, the wave field calculations move from the physical domain to the calculation domain. The mapping method deals with the irregular surface and the low-velocity layer underneath it using a fine grid. For the deeper high-velocity layers, the use of a fine grid causes local oversampling. In addition, when the irregular surface is transformed to horizontal, the flattened interface below the surface is transformed to curved, which produces inaccurate modeling results because of the presence of ladder-like burrs in the simulated seismic wave. Thus, we propose the mapping method based on the dual-variable finite-difference staggered grid. The proposed method uses different size grid spacings in different regions and locally variable time steps to match the size variability of grid spacings. Numerical examples suggest that the proposed method requires less memory storage capacity and improves the computational efficiency compared with forward modeling methods based on the conventional grid.
