【论文阅读】9-Photo Tourism: Exploring Photo Collections in 3D

1、论文简介：

论文主要针对大场景、大数据量，快速建模，同时恢复相机位姿。
进一步地，构建相应的工程项目（Photo Tourism）

2、核心实现-采用incrementalSFM

2.1、First single pair

Matches 最多，large baseline

2.2、Add one camera

2.2.1 条件

被估计的track 数目最多,M

2.2.2 initialize

1. extrinsic & intrinsic ：DLT (direct linear transform) + RANSANC( extrinsic & intrinsic)
2. 3D landmarks: triangularization(Estimated cameras + added camera )

2.2.3 Optimize:

1. 优化对象：Estimated cameras + added camera + estimated tracks + unknown tracks
2. Unknown tracks 加入到优化计算的条件：

• The number of the track is observed by the estimated camera >=1
• 三角化的结果满足精度

Trick:

BA优化过程中，每一轮迭代后，remove outlier tracks

2.2.4 add camera 的具体实现

Add one camera ---add two camera M + any 0.75M (the number of observed tracks)

3、Geo-registration ---- user interaction

Relative location ----similarity transform -------absolute location

4、Scene align to DEM

Rerun the sfm BA:  object function --- to fit the correspondences between the point cloud and DEM

5、其他：未看