1. 要解决的问题：场景重建 -> colmap效果很好，具体方法 -> colmap在地面、墙面这些无纹理区域效果不好，具体原因。
2. 传统方法：用plane帮助场景重建 -> 流程复杂，很多要调的参数 -> plane效果不好，重建质量不好。
3. 最近的方法：nerf、volsdf、neus在物体重建上效果很好 -> experiments show that 他们在室内无纹理区域效果不好 -> 大片无纹理区域存在很多可以解释的几何。
4. 我们的方法：用语义帮助重建，并且在重建的同时优化语义信息 -> Specifically, 检测地面和墙面，让他们符合语义性质 -> 假设曼哈顿结构，地面和墙面上的表面点normal要符合相应的性质。墙面的normal是优化出来的。-> 考虑到分割可能不准的问题，我们定义了semantic mlp。-> 利用multi-view consistency提升semantic segmentation的准确性。同时用几何的loss优化分割的probability。
5. Experiments
要解决的问题
1. Reconstructing 3D scenes from multi-view images is a cornerstone of many applications such as augmented reality, robotics, and autonomous driving.
2. Given input images, traditional methods generally estimate the depth map for each image based on the multi-view stereo and then fuse estimated depth maps into 3D geometries.
3. Although these methods achieve impressive reconstruction results, they have difficulty in handling low-textured regions such as floors and walls of indoor scenes due to unreliable matching in these regions.
传统方法利用planar prior来提升效果
1. To overcome this problem, some methods utilize the planar prior to help reconstruction.
2. They 怎么用planar prior的
3. plane怎么建模：triangulation \cite{planar prior}, superpixel \cite{tapa}, or learning-based plane segmentation methods \cite{}.