OpenMVS Texture code reading

• TextureMesh
  • FaceViewSelection
    • ListVertexFaces✔️ // extract array of triangles incident to each vertex
      • EmptyExtra✔️
      • ListIncidenteFaces✔️
      • ListBoundaryVertices✔️, // 边界点检查，在面中，只使用了1次的点是边界点，另外，代码中假设每个点通常不会超过12个面使用
        • 
    • // create texture patches✔️
      • ListCameraFaces // list all views for each face
        • //create vertices octree of vertices✔️
        • //extract array of faces viewed by each image✔️
          • //compute gradient magnitude✔️
          • //select faces inside view frustum✔️
            • plane with AABB✔️
          • // project all triangles in this view and keep the closest ones
            • // skip face if not completely inside✔️
              • using TransformPointW2C and TransformPointC2I to check the point is in image plane, OpenMVS use 3 pixel as border width ✔️
            • // skip face if the (cos) angle between the view to face vector and the view direction is negative✔️
              • (faceCenter.z <= REAL(0))
              • faceCenter.dot(normalPlane) >= ZEROTOLERANCE<REAL>()
            • // draw triangle and for each pixel compute depth as the ray intersection with the plane✔️
              • Advanced Rasterization by Nick (Nicolas Capens) 2004
              • Advanced Rasterization🤔
              • libs/Common/Types.inl:2474/TImage<TYPE>::RasterizeTriangle
              • OpenMVS: Using block size eq 8, and Half-Space
          • // compute the projection area of visible faces
            • 遍历照片上的像素，获取最近的face最近的view, 计算face颜色均值✔️
          • (fOutlierThreshold > 0) // try to detect outlier views for each face
            • 取face 3个通道的均色，对均色做多元正太分布计算，如果偏差大于阈值，则视为outlier✔️
              • LU分解取逆
      • // create faces graph✔️
        • 遍历face以及面的邻接faceAdj, 若一个可见，一个不可见，则增加seam
      • // assign the best view to each face✔️
        • // find connected components✔️
          • 根据前面计算的face邻接关系计算连通图
        • // map face ID from global to component space✔️
          • 为每个面计算连通图的新id，component1: [1,2,3,...,c1], component2: [1,2,3,...,c2]
        • // normalize quality values✔️
          • 计算图片质量[二阶导积分]的95分位置近似值
        • // initialize inference structures✔️
          • numNodes=连通图face的数量，如果数量小于等于1，则忽略
          • SmoothCost=0[if node eq]~1000
          • Neighbors
        • // set data costs✔️
          • // set costs for label 0 (undefined)✔️
            • dataCost = fRatioDataSmoothness * 1000 for min of all labels
          • // set data costs for all labels (except label 0 - undefined)✔️
            • dataCost = $\big(1-\max(1, \frac{quality}{quality_{p95}})\big)*1000$
          • Label = 0, 未定义，label >= 1是指其他图片✔️
        • // assign the optimal view (label) to each face✔️
          • Optimize 优化单位是连通图✔️
            • ComputeEnergy✔️
              • $\sum{cost_{data}} + \sum(cost_{edge})$
              • $cost_{edge} = 1000$ if label is not equal
            • Optimize(n=1)✔️
              • for(n times)
                • 计算每个面在选择不同label时得到的minEnergy
                • Swap newMsgs and oldMsgs, oldMsgs - min(oldMsgs)
              • 更新选择的label和dataCost✔️
            • 根据能量差分提前返回✔️
        • // extract resulting labeling✔️
          • label-1，调整索引
      • // create texture patches✔️
        • // divide graph in sub-graphs of connected faces having the same label✔️
          • 计算图片不一致的边缝A，对于相邻一侧无label的也作为边缝B
          • 去掉A的连通边
        • // find connected components: texture patches✔️
          • 计算新的连通图
        • // last texture patch contains all faces with no texture✔️
          • 每个Texture 收集每个连通体的face
        • // remove all patches with invalid label (except the last one) and create the map from the old index to the new one✔️
  • GenerateTexture
    • // project patches in the corresponding view and compute texture-coordinates and bounding-box
    • // project vertices and compute bounding-box✔️
      • 对每个patch的面映射到图片中后，算一个2d的aabb
    • // compute relative texture coordinates✔️
      • 计算offset，修改相应的坐标
    • // init last patch to point to a small uniform color patch✔️
      • 默认纹理位置
    • // perform seam leveling
      • // create seam vertices and edges✔️
      • // perform global seam leveling🤔
        • 利用ConjugateGradient求解Tikhonov's Gamma matrix使得全局颜色差分较小的新解
        • // find the patch ID for each vertex✔️
          • 对点赋值patch_id，边缝点使用idxSeamVertex
        • // assign a row index within the solution vector x to each vertex/patch✔️
          • 生成一个patch数组vertpatch2rows, 行号为点索引，二阶索引为patch，对于单个patch的点，只赋值一个patch，对于边缝点多个patch的点，赋多个patch，值为递增值rowsX
        • // fill Tikhonov’s Gamma matrix (regularization constraints)✔️
      • // perform local seam leveling✔️
        • 优化patch间的线条颜色
    • // merge texture patches with overlapping rectangles
      • // translate texture coordinates
      • // join faces lists
      • // remove the small patch
    • // create texture
      • // arrange texture patches to fit the smallest possible texture image
      • // increase texture size till all patches fit
    • // create texture image
      • // copy patch image
        • // flip patch and texture-coordinates
      • // compute final texture coordinates
        • // translate, normalize and flip Y axis