July

1st week

• Documentation:

  • Added a new section for computing depth map using stereo matching
  • Studied the method of depth map fusion using a cost volume of a regular voxel grid
  • Added new details for minimizing the first-order primal dual algorithm from @chambolle2011first

• Implementation

  • Tested Gallup implementation of local stereo matching using SAD in CUDA

2nd week

• Studying Dense Tracking and Mapping (DTAM) implementation by @newcombe2010dtam
  • Base mesh is build over sparse point cloud from PTAM by the polygonalization of the function's zero level set of a function $f$ fit to the data points $\mathbb{R}^3 \rightarrow \mathbb{R}, f(x) = 0$
  • Model predictive optical flow using projected images from the base model
  • Use of dense TVL1 optical flow implementation to track pixels
  • Mesh optimization using scene flow to compute vertex updates $\triangle \mathbf{x}_j$ which lie in the ray from the reference camera intersecting the base model vertex $\mathbf{x}_j$

3rd week

• Added documentation about Scene Flow and Optical Flow by @vedula1999sceneflow
• Put some glue to all ideas making a new introduction to the main framework based on vector spaces over fibrations

4th week

• Testing feature tracking based on FAST detector combined with fast matching using ZNCC search
• ORB and FAST feature detectors are tested in the GPU. Both are fast and more stable than the goodfeaturestotrack

5th week

• Removed submodule dependencies from OpenCV and OpenMVG: they increases project complexity and at currently they are not modified
  • In fact, QtCreator does not recognize the OpenCV INCLUDEPATH using submodules
• Debugging triangulation procedures to reduce the high Mean Reprojection Error (from 1000 to less than 1.0!)

References