go>> pais-mvs>> 返回
项目作者： adahbingee

项目描述 ：
Multi-view stereo image-based 3D reconstruction
高级语言： C++
项目地址: git://github.com/adahbingee/pais-mvs.git
创建时间： 2015-03-13T01:13:52Z
项目社区：https://github.com/adahbingee/pais-mvs
开源协议：
下载


pais-mvs

Multi-view stereo image-based 3D reconstruction

PAIS Multi-view Stereo

image

image

image

library:

OpenCV 2.4.2

Point Cloud Library 1.6.0

Publications

  • “A novel 3D dense reconstruction with high accuracy and completeness”. Zen Chen, Wen-Chao Chen, And Ping-Yi Sung. IEEE International Conference on Multimedia and Expo Workshops ICMEW 2013

  • “Stochastic Optimization Based 3D Dense Reconstruction from Multiple Views with High Accuracy and Completeness”. Wen-Chao Chen, Zen Chen, And Ping-Yi Sung. Journal of Information Science And Engineering 31, 131-146, 2015

  • “Stocastic Optimization Based 3D Reconstruction with High Accuracy and Completeness”. Wen-Chao Chen, Ping-Yi Sung And Zen Chen . The 25th IPPR Conference on Computer Vision, Graphics, and Image Processing, CVGIP 2012 Excellent Paper Award

Link

Pre-build binary

Introduction

  1. Input / Output file format
  2. Config parameters
  3. Command line instruction

Input file format

nvm file format

NVM file is come from VisualSFM data structure. Here is the example of the data structure. The detail can be found in FileLoader.

You can skip the calibration step and just use the Structure from Motion result output from VisualSFM.

  1. NVM_V3 
  3. 5
  4. pawn0013.jpg    614.095397949 0.705410371683 0.160690743319 0.671401589359 0.160605237544 -0.556085150075 0.0481223921551 -0.00781510757143 -0.199289312888 0 
  5. pawn0010.jpg    616.175537109 0.90353903514 0.221746421078 0.3576944596 0.0806247263945 -0.880841878288 0.0327703491031 -0.684201024844 -0.209314043486 0 
  6. pawn0011.jpg    612.03302002 0.85241383667 0.2037593266 0.469072019941 0.108830220502 -0.71971232163 0.0433857776889 -0.492035476323 -0.207263977174 0 
  7. pawn0012.jpg    611.360473633 0.786507583571 0.183363764635 0.573952646995 0.135504187104 -0.608685012281 0.0487066227347 -0.263440114899 -0.203210786458 0 
  8. pawn0014.jpg    617.585876465 0.611485687162 0.135944898976 0.757586998462 0.183482834469 -0.572254659063 0.0434025057556 0.255716172724 -0.198563271584 0

The first row shows the NVM file version.

Third line 5 indicate there are 5 cameras in the following list.

Each list entry:

  • image file name, can be PNG, JPG, BMP, or any OpenCV supported image format. Note that the file path can be relative or absolute path.
  • focal length
  • 4 elements vector (k wx wy wz) indicate the quaternion rotation. Transformation form quaternion to 3x3 rotation matrix can be found in code. Camera.
  • 3 elements vector (x y z) indicate camera center in world coordinate.
  • radial distortion parameter. The usage of parameter can be found in camera projection.
  • 0 for end of line

nvm2 file format

In the NVM file format, the camera principle point is assumed at the image center, and can’t to be edit.

NVM2 file specify the principle point(px, py) after focal length. Others the same.

mvs file format

MVS file is saved in a binary way for internal use, for example: MVS viewer, patch filtering.

Output file format

mvs file format

MVS file is saved in a binary way for internal use, for example: MVS viewer, patch filtering.

ply file format

ply is the point cloud model. Can be open using MeshLab.

psr file format

psr file is ready for the Poisson surface reconstruction.

Config parameters

config.txt should be placed in the same directory of main program.

  1. ### patch optimization configuration ###
  2. # patch radius (patch size = 2*radius+1)
  3. # !!!!check distWeighting if patchRadius have been changed!!!!
  4. # default 15
  5. patchRadius            15
  7. # reduce normal search range for expansion patch (reduce to +- pi/reduceNormalRange)
  8. # 1 stands for search 360 degree, 2 for 180 degree
  9. # smaller for larger cell size
  10. # default 2 (reduce to 180 degree)
  11. reduceNormalRange    2
  13. # enable or disable adaptive distance weighting (0: disable, 1: enable)
  14. # default 1
  15. adaptiveDistanceEnable        1
  16. # distance weighting to patch center
  17. # default(patch radius/3) 5
  18. distWeighting        5
  20. # enable or disable adaptive difference weighting (0: disable, 1: enable)
  21. # default 1
  22. adaptiveDifferenceEnable    1
  23. # difference weighting
  24. # default(128*128) 16384
  25. diffWeighting        16384
  27. # enable or disable adaptive gradient weighting (0: disable, 1: enable)
  28. # default 0
  29. adaptiveGradientEnable        0
  30. # gradient maginitude weighting
  31. # default 10.0
  32. gradientWeighting    10.0
  34. # visible camera angle correlation
  35. # smaller for sparse camera deploy
  36. # default 0.87 radian for 30 degree
  37. visibleCorrelation    0.7
  39. # depth search range scalar
  40. # smaller for smoother object, larger for complex/sharp object
  41. # smaller for small cell size, larger for big cell size
  42. # larger depthRangeScalar cloud avoid local trap!
  43. # beta in paper
  44. # default(cellsize/2) 2
  45. depthRangeScalar    8
  47. ### PSO optimization configuration ###
  48. # particle number
  49. # default 30
  50. particleNum            15
  51. # maximum iteration number
  52. # default 60
  53. maxIteration        30
  55. ### expansion configuration ###
  56. # image cell size, control the density of patches
  57. # default 4
  58. cellSize            2
  59. # maximum patch number in single cell
  60. # default 3
  61. maxCellPatchNum        3
  62. # patch expansion strategy (Best: 0, Worst: 1, Breath: 2, Depth: 3)
  63. # default 0
  64. expansionStrategy    0
  66. ### level of detail configuration ###
  67. # texture variation
  68. # default 36
  69. textureVariation    36
  70. # minimum level of detail
  71. # default 0
  72. minLOD                0
  73. # maximum level of detail
  74. # default 15
  75. maxLOD                15
  76. # level of detail ratio
  77. # default 0.8
  78. lodRatio            0.8
  80. ### filtering configuration ###
  81. # minimum visible camera number
  82. # default 3
  83. minCamNum            3
  84. # minimum texture correlation
  85. # default 0.9
  86. minCorrelation        0.9
  87. # minimum patch homography ellipse axes ratio
  88. # default 0.15
  89. minRegionRatio        0.15
  90. # maximum fitness (critical effect the completeness)
  91. # default 10.0
  92. maxFitness            10.0
  93. # neighbor patch search radius scalar (from TVCG09 PCMVS)
  94. # the neighbor radius = pow(volume, 1.0/3.0) * neighborRadiusScalar
  95. # default 0.0025*(cellsize*2)
  96. neighborRadiusScalar        0.01

Command

There are 4 commands for our program.
TMVS.exe -r
TMVS.exe -f
TMVS.exe -v
TMVS.exe -a

Reconstruction TMVS.exe -r

Run multi-view stereo reconstruction.

The program will use the N-view matching points as the seed points from VisualSFM.

When there are no seed points, it will run N-view matching algorithm by SIFT feature and perform N-view matching with epipolar constrain.

Note that the config.txt must be initialized.

  1. TMVS.exe -r input.nvm

There are 5 output files from reconstruction command.

init.mvs the input seed points from VisualSFM.

seed.mvs seed points after optimization.

exp.mvs expansion point cloud.

exp.ply expansion point cloud.

exp.psr expansion point cloud.

Patch post-processing filtering TMVS.exe -f

Run post-processing filtering after reconstruction. Note the filtering process only accept mvs file.

  1. TMVS.exe -f exp.mvs

MVS Viewer TMVS.exe -v

Rendering result in PCL framework.

press h to see help function.

  1. TMVS.exe -v exp.mvs

MVS Viewer TMVS.exe -a

Rendering result in PCL framework with expansion animation.

press h to see help function.

  1. TMVS.exe -a exp.mvs