多视图几何轻量级三维重建算法

Lightweight 3D reconstruction algorithm based on multi-view stereo

针对现有深度学习三维重建网络内存消耗严重、效率低下的问题,提出了高效的多视图几何三维重建网络(high efficiency multi-view stereo network,H-MVSNet)模型,将原始图片序列和预测的粗略深度图融合,进一步提高最终深度图的质量;构建轻量级的特征提取模块和正则化模块,减少提取冗余度;采用由粗到精的策略,建立高效的深度图细化模块,减少计算量。实验表明,H-MVSNet模型在DTU数据集中的精度误差可达0.327 mm,计算一张分辨率为640×480的深度图仅需0.44 s,内存消耗可低至2.46 GB,显著提高了三维重建的精度和准确度。

In order to solve the problems of severe memory consumption and low efficiency in existing 3D reconstruction networks based on deep learning,we propose a lightweight multi-view stereo 3D reconstruction algorithm,which fuses the original image sequence with the predicted rough depth map to further improve the quality of the final depth map.A lightweight feature extraction module and regularization module is constructed to reduce the redundancy of extraction.The strategy from coarse to fine is adopted to establish an efficient depth map refinement module to reduce the amount of calculation.Experiments show that the accuracy error of H-MVSNet proposed in this study in the DTU dataset reaches 0.327 mm.It only takes 0.44 s to calculate a depth map with a resolution of 640×480,and the memory consumption is as low as 2.46 GB,which significantly improves the accuracy and accuracy of 3D reconstruction.