danini / graph-cut-ransac

The Graph-Cut RANSAC algorithm proposed in paper: Daniel Barath and Jiri Matas; Graph-Cut RANSAC, Conference on Computer Vision and Pattern Recognition, 2018. It is available at http://openaccess.thecvf.com/content_cvpr_2018/papers/Barath_Graph-Cut_RANSAC_CVPR_2018_paper.pdf

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Support float/double image width and height

Parskatt opened this issue · comments

Hi,

Currently the python wrapper, see e.g. here:

int findFundamentalMatrix_(
// The 2D-2D point correspondences
std::vector<double>& correspondences,
// The probabilities for each 3D-3D point correspondence if available
std::vector<double> &point_probabilities,
// Output: the found inliers
std::vector<bool>& inliers,
// Output: the found 6D pose
std::vector<double> &fundamental_matrix,
// The images' sizes
int h1, int w1, int h2, int w2,

only supports integer image sizes.

As far as I can tell, these are always cast to doubles, so it would make sense to support arbitrary numbers.

In my use case I have modified the intrinsics, and hence the "virtual" image sizes become doubles. I'm unsure if rounding these to integers pose any issues to the RANSAC loop, so would prefer to insert the exact values.

Although reading the code it seems these only control neighbourhood sizes so I guess rounding shouldnt be a big issue.

Best,
Johan

Rounding should not cause any problem I think. The sizes are only used for specific samplers and for the neighborhood calculation.