Traditional reconstruction methods use photometric consistency to compute dense 3D information. Although these methods are in ideal Lambertian scenarios, the accuracy is already very high.
But traditional limitations,例如弱纹理，高反光和重复纹理等，使得重建困难或重建的结果不完整 .
A learning-based approach can 引入比如镜面先验和反射先验等全局语义信息，使匹配更加鲁棒，从而解决传统方法无法克服的难题.
It is difficult for traditional vision algorithms to make new breakthroughs. Various fields are researching in the direction of deep learning. In recent years, the papers of major vision conferences are based on deep learning to achieve 3D reconstruction. The recruitment of various major companies is also paying more and more attention to deep learning. 3D reconstruction method.
Article directory
1 Tutorial directory
A total of 50 pdf tutorial files,
It mainly focuses on the theoretical and practical explanations of PatchMatchNet and MVSNet, supplemented by MVS explanations of other deep learning series.
The main content includes:
explanations and practical operations of 3D reconstruction principles based on deep learning;
mainstream deep learning MVS network series;
unsupervised MVS, semi-supervised MVS, self-supervised MVS explanations;
pytorch, neural network and other basic necessary knowledge explanations;
visualization tools ;
