MVS学习（二）：MVS重建的数据获取方法推荐

MVS学习（一）：综述论文Multi-View Stereo: A Tutorial阅读记录
Image acquisition is the first critical step for successful MVS。
无论是三维重建任务，还是其他各种cv相关任务， 图像获取的质量往往决定功能效果的上限，也是对效果影响最大的因素，但是实际执行过程中往往容易被忽略，或者不知道从哪些角度考虑去优化图像质量，下文说明针对mvs任务，如何获取符合要求的高质量图像。

相机模型的精度

Accuracy of the camera models: MVS techniques are highly dependent on the accuracy of the camera parameters. Typical reprojection error RMSE should be sub-pixel, ideally smaller than 0.5 pixels. In case reprojection error is large, one possibility is to shrink the input images and modify the corresponding camera parameters, which will reduce the reprojection error proportionally to the shrinkage ratio

这里主要说明相机参数对mvs结果的影响。mvs之所以目前效果有很多提升很大的愿意就是以sfm为代表的相机参数获取方法质量的提升，所以如何获取更高精度的相机参数是优化最终mvs效果的一个思考方向。

图像分辨率

High resolution images bring up details that can be used to uniquely identify a pixel from
its neighbors, thus improving the correspondence cue used by MVS algorithms to find similar pixels across multiple images

这里说明了为什么像素点多对重建效果有提升：更多的像素点有利于发现与邻近像素点之间的差异（低解析度可以理解成在高解析度图片的下采样，细节信息肯定有损失），细节信息丰富有利于重建时候在多张图像中寻找定位相似的像素点（高分辨率的影响也可以理解成纹理信息更丰富）

Note however that by resolution we do not mean just having lots of pixels, the quality of camera lens also matters. Having a very high-res image captured with a poor quality lens will not improve results, and it may actually make them worse, e.g. due to worse results at the SfM stage.

分辨率有一个点值得注意，即并不是分辨率越高（像素点越多）最终的重建效果越好。相机的镜头质量更重要，相机镜头的畸变等会降低sfm的效果，单纯提升分辨率而镜头质量差会导致更差的结果，sfm对最终效果的影响巨大。

图像重叠度

For MVS to work correctly, multiple images need to see the same piece of geometry from multiple view points

图像数量

图像数量越过通常效果越好，但是很多实际场景应用必须在图像数量和分辨率之间做平衡

However, if one has
to choose between image resolution and number of images, there is no easy decision. MVS algorithms reconstruct more details from higher resolution images, as MVS suffers little from ambiguous matches. On the other hand, high resolution images become an issue for SfM, as the so-called ratio-test would reject many feature matches. Therefore, if good camera calibration is available, in general one should choose image resolution over number of images

更多的图像数量对mvs基本都是积极影响，但是可能会导致sfm算法的精度下降，所以，当分辨率和图像数量必须权衡的时候，优先保证分辨率。

图像质量

Although there exist many algorithms that are robust to illumination variations across the images, the more stable it can be,the better. For example, flash changes shading and shadowing effects in every image, and should not be used for weakly textured surfaces

尽管算法对光照变化有一定的鲁棒性，但是如果可以保证光照稳定还是要尽量保证。光照越稳定，通常结果越好。尤其对于贫纹理表面物体更要保证光照的稳定性。

Ideally all，the images used in MVS should be all-in-focus. This can be achieved
by using small apertures and large exposures, within the limits of a particular application.

理论上，用于重建的图片要求全对焦，也就是物体要处于相机的焦点上。这对于小物体还好保证，大场景如何做也有很多实际的改进。
MVS学习（一）：综述论文Multi-View Stereo: A Tutorial阅读记录