Pierre Chatelier

> may i try solving this bug? As far as I am concerned you are welcome. I wonder if the "best" result will be considered the one from cuda::remap() or...

If I remember correctly, in early CUDA (older compute compatibilities), reading from a texture was a huge performance gain compared to (even read only) global memory. Thus, using a texture...

I am not familiar at all with CMake, I don't think I can submit a PR. ~Would you like me to send you the CMakeCache.txt of my zlib, HDF5, and...

@jmckenna I just cloned https://github.com/jmckenna/netcdf-c.git and built it with my usual Windows+CMakeGUI+VS2019 target I confirm that once generated, the Visual Studio project does not require any more manual modification to...

I investigated a little and found some clues for this error-case of that particular set of points P. - First, the result looks correct when feeding the algorithm with the...

Older issues related to [fitEllipse()](https://github.com/opencv/opencv/blob/4.x/modules/imgproc/src/shapedescr.cpp#L504) : #4515, #6544 [https://stackoverflow.com/questions/34176179/opencv-ellipse-fitting-extract-parameters](https://stackoverflow.com/questions/34176179/opencv-ellipse-fitting-extract-parameters) The related paper : [rp810.pdf](https://homepages.inf.ed.ac.uk/rbf/MY_DAI_OLD_FTP/rp810.pdf) [ellipse-pami.pdf](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/ellipse-pami.pdf)

I wonder if the problem comes from an inherent "flaw" or "unstability" in the space of non-unique solutions when the SVD is facing over-determined data. I came across this paper...

I confirm that changing the SVD algorithm will change the results significantly. I have implemented the "spherically normalized" SVD mentioned in the paper, and replaced the first `SVDecomp()` of `fitEllipseNoDirect()`...

> " First, the result looks correct when feeding the algorithm with the convex hull of P (as mentioned by the OP)" i think adding `convexHull` in the `fitEllipse` source...

I have a new interesting remark. Let P be the "faulty" input set of Points used from the beginning of this discussion. I just assume that the clusters of points...