【论文阅读】18-Large-scale and high-quality Multi-view stereo
0 basic Information
TY - JOUR
AU - Vu, Hoang
PY - 2011/12/05
SP -
T1 - Large-scale and high-quality multi-view stereo
ER -
1、Photometric robust variational refinement
1.1、 Refine的必要性
Initial surface M0:
- Isolated T (基于可视性约束)
- big size T(lack points–textureless)
解决办法:
- remove T (size // number)
- Landscape (far from camera – inexact calibration)—cut (无法refine)
1.2、 refine 核心
variational refinement ---optimize photo-consistency
构造能量函数
E=(E_error+ E_fair + regularize)
E与surface(M)有关,对E求最小值,以M0(initial surface)为迭代初值
求解方法: graph-cut
1.3、E-能量函数构造
1.3.1、E_error --基于 E_photo 改进
H(Ii,Iij) ---- photo-consistency –常用: -NCC
1.3.2、 E_fair
- 作用:measure the total curvation –surface 测度
- 数学表达: thin-plate (薄板样条函数)
- 简化:bi-Laplacian
1.3.3、Regularize
作用:smoothing terms (常用思路)
Twofold:
- Photo-consistency –reliable (textureless) 特别是在弱纹理区域,给予photo-consistency 一定的置信度 数学改进: 引入r( xi) : reliability factor –Pair( I,J)
- 引入scale factor,使得E_error & E_fair homogenized
Scale factor=【avg(depth) /focal length】^2
1.4、优化求解过程
重点
:surface 离散化–Surface – T –vertexes
PS:两种思路: 先优化,再离散 // 先离散,再优化 (本文采用此方式)
离散化后,对E的梯度计算的推导:
E_error 关于3D point X 的梯度的几何意义
: 2D pix 对应3D point–投影点在 该 X邻域的所有T内,对这样的2D pix 加权
权重(重心坐标):
1.5、 mesh resolution:Adaptively: subdivide
Camera pair(I , J)
只要其中一个相机满足以下条件,T subdivide条件:T 重投影到图像对应的像素数量超过阈值(本文设置为16
Subdivide 方式:
one-to –four – edge middle point
Loop C T. Smooth Subdivision Surfaces Based on Triangles[J]. Department of Mathematics the University of Utah Masters Thesis, 1987
1.6、 结果及评估
Accuracy --- 针对refienment
算法产生的mesh与gt计算准确度(基于vertex or face),精确度在前90%的vertex or face 对应的误差
(空间距离?! 以mm为单位)
Completeness --- 针对initial surface
以1.25mm为阈值计算refined mesh 相对于gt的完整度
1.7 其他
Level set — surface 重建的一种方法,level set 等值线 – 隐函数类型
2、Surface triangular mesh merging
2.1、pre
- 前提:sub mesh –same coordinate
- Sub mesh merge –存在double & gap
2.2、step1–overlap detection
- T(triangle)–sphere
- 球心 : 质心G
- 半径:
- collision detection
利用kd-tree 进行collision detection:
Different leaves(叶节点): not intersection
Same leaves(叶节点): intersection
2.3、step2 --CDT (constraint DT)
- pre:surface 划分:
- Overlap
- Overlap区域的vertex– ring face in non-over ---- F
- Non-overlap
- Edges(区分 non-overlap && F) –—E
- 参与DT构建的elements:
- Overlap vertex:
重要!!! 先对position 接近的vertex进行合并,去除冗余
Position接近—sphere(r = ε(10^-6)/2) collision detection(kd -tree 同上)- F :error(T intersection) detection && remove
- Infinite vertex — open surface
- CDT的构建方式:
F中face保留或者以某种方式细分,未细究。论文实现方式:software --Si Hang’s software tetgen
2.4、step3 --graph & minimize cut
- Dual graph ( 对偶图的作用:采用s-t的最短路径规划来实现最大流、最小能量函数的问题)
- 能量函数:
重点!!!
Graph – weight 设置(与主流思想不一致,基于先验约束!!!)
1) Constraint cost
The face of f or sub_f ( F) 邻接DT:
- DTi – Front of the face –outside – si = si + Wc
- DTj – behind of the face –inside – tj = tji + Wc
- Wc : big value
原理
: 具有较大置信度的 inside / outside 的先验,可通过增大对应cost,使得最终mini-cut 与先验尽可能保持一致
实质
:由sub mesh 已经构建的F (ring face in non-over),在最终的surface中也有较大可能存在,因此基于F的DT inside (behind)/ outside(front) 具有较强置信度
作用
:Constrained facets with right orientation appear on the final surface
2) Normal cost
Mesh中vertex normal vector 指向outside,基于该先验,设置s/t cost 对于某vertex,计算
normal vector N, 其邻接DT
- Finite DT (包含N ) + infinite DT —outside si = si + Wn
- FiniteDT(包含-N) —inside tj = tj + Wn
作用
:The normal costs encourage the surface to pass the soup of facets in overlapped areas.
3) Quality cost
!!!! DT-DT weight
,与surface quality 有关,因此,只有wij_包含可视一致性约束 + surface
quality 具体细节同参考论文
25、step4–S1 优化
2.5.1、S1-S2
Steiner point — sub triangles (未细究)
定义见链接🔗
2.5.2、S2-S3
!!!再次运用graph-cut
:
- Node: S2的triangle 三角面片
- Edge:the sharing edge
- Weight – 基于先验【F && E】的s/t weight 改进:
- Triangle in F – “inside” — (表示在最终的surface内部,与先前的inside含义不同)
- Triangle adjacent to the E — “outside”------利用的border信息
2.5.3、S3-S4
- 优化原因: 基于manifold constraint 【vertex && edge的拓扑邻接关系】
- 优化方式:vertexes in S3 + vertexes in S3’s ring face (non-over) —then?? (不清楚)
2.6、step5 --merge
2.7、final mesh—Laplacian smoothness
2.8、result
- 数据集
City-scene dataset:different coordinates –overlap zone: sub mesh intersection inside the box — accelerate the overlap detection
- 参数设置
- 针对 far vertices,效果不好—解决办法:
- triangle size 阈值
- graph-cut 过程中对big triangle set strong penalty
- Device info
3、other
- PLS( piecewise linear system) —见论文附录 130/153
- 5.3 Merging meshes from partition of bounding box --未看