[GSoC] Point Cloud Object Fitting #2584
Conversation
| #include <vector> | ||
| #include <utility> | ||
| #include "opencv2/viz.hpp" | ||
| #define PLANE_MODEL 1 |
There was a problem hiding this comment.
enums are a better choice than defines
|
@devanshbatra04 Any update on this, please? |
|
|
||
| bool getSphereFromPoints(const Vec3f* &points, const std::vector<unsigned> &inliers, Point3d& center, double& radius) { | ||
| // assert that size of points is 3. | ||
| Mat temp(5,5,CV_32FC1); |
There was a problem hiding this comment.
this should be a 4x4 Mat, with the last column set to 1
(as it is now, it has 10 uninitialized values, wrecking any determinant calculated later)
|
|
||
| assert(normals_cld.cols == cloud.cols); | ||
|
|
||
| const Point3d* points = cloud.ptr<Point3d>(0); |
There was a problem hiding this comment.
this must be Point3f (the point cloud is in float format, not double)
|
|
||
| double magnitude_abc = sqrt(ModelCoefficients[0]*ModelCoefficients[0] + ModelCoefficients[1]* ModelCoefficients[1] + ModelCoefficients[2] * ModelCoefficients[2]); | ||
|
|
||
| assert (magnitude_abc == 0); |
| return result; | ||
| } | ||
|
|
||
| viz::WCylinder SACCylinderModel::WindowWidget () { |
There was a problem hiding this comment.
probably, we should outsource any drawing functionality to the samples.
having a dependancy of the library code to the viz module (which again depends on vtk) is some major bummer, this could/should be made optional
| break; | ||
| } | ||
| vector<unsigned> latest_model_inliers = inliers.back(); | ||
| num_segmented_points += latest_model_inliers.size(); |
There was a problem hiding this comment.
bail out, if num_segmented_points==0 ?
(since the rng is reseeded inside fit_once() it cannot do anything different, if the labels array wasn't changed, leading to an infinite loop)
| } | ||
|
|
||
| if (model_type == CYLINDER_MODEL) { | ||
| assert(this->normals_available == true); |
There was a problem hiding this comment.
so, currently, normals are not optional here.
maybe we can add another method for adding a hypothesis based on 3 points (2 points parallel to the axis, just on the outer hull, and a 3rd, somewhat opposite) ?
generating normals from 12 (!) knn neighbours only seems to work nicely for very dense clouds, but not for sparse, synthetic or sfm generated clouds
|
Hi, @devanshbatra04 it's been quite some time without an update. Can you please refactor the code as suggested so we can merge the PR? |
|
would there be a way to work on this (and still keep the original commits / @devanshbatra04 as author) ? i think i got some improvements |
Hi, we notice that you are improving the object detection feature. Our group hopes to complete the feature. What do you think we can do for this feature? Thank you. |
|
if it is about the gsoc idea, |
We really don't know that this project has been listed as GSoC project again. We will try to apply. Thank you. |
|
jenkins cn please retry a build |
|
Is this pull request superseded by opencv/opencv#21276 ? |
This PR would contain the work done towards the GSoC 2020 project on Point Cloud object fitting.
This is a test to see whether the PR builds. I will continue adding changes.
cc: @mihaibujanca
Pull Request Readiness Checklist
See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request
Patch to opencv_extra has the same branch name.