aboulch / normals_hough Goto Github PK

The Python wrapper does not compile under Windows. In this fork I added some quick and dirty fixes you could maybe use.

The issues I encountered:

• setup.py does not include the numpy's include directory.
• The compile flags specified in setup.py only apply to GCC, not to MSVC.
• MSVC's math.h does not define M_PI unless _USE_MATH_DEFINES is defined.
• The typedef kd_tree uses the matrix type Eigen::MatrixX3d. This results in the KDTreeEigenMatrixAdaptor's IndexType to be __int64 under MSVC. In MSVC, long int cannot be converted to __int64 and resulted in type errors at places where knnSearch() and list_of_triplets() were called. I solved this by introducing a typedef pointIdxType which is __int64 when using MSVC and long int otherwise.

Below is the source of list_of_triplets from normals.h:

	/*!
	 * generates a list of triplets
	 * @param triplets : table of 3-vector to fill with the indexes of the points
	 * @param number_of_points : number of points to consider
	 * @param triple_number : number of triplets to generate
	 * @param vecRandInt : table of random int
	 */
	inline void list_of_triplets(Eigen::MatrixX3i &triplets,
			const int &number_of_points,
			const unsigned int &triplet_number,
			std::vector<int> &vecRandInt){

		unsigned int S = vecRandInt.size();
		triplets.resize(triplet_number,3);
		unsigned int pos=vecRandInt[0]%S;
		for(unsigned int i=0; i<triplet_number; i++){
			do{
				triplets(i,0) = vecRandInt[pos%S]%number_of_points;
				triplets(i,1) = vecRandInt[(pos+vecRandInt[(pos+1)%S])%S]%number_of_points;
				triplets(i,2) = vecRandInt[(pos+vecRandInt[(pos+1+vecRandInt[(pos+2)%S])%S])%S]%number_of_points;
				pos+=vecRandInt[(pos+3)%S]%S;
			}while(triplets(i,0)==triplets(i,1) || triplets(i,1)==triplets(i,2) || triplets(i,2)==triplets(i,0));
		}
	}

The do-while-loop may run forever. This could happen if vecRandInt[(pos+3)%S]%S evaluates to 0 and the condition (triplets(i,0)==triplets(i,1) || triplets(i,1)==triplets(i,2) || triplets(i,2)==triplets(i,0)) holds true.

when compiling on ubuntu 16.04, the test_cpp fails with

........../normals_Hough/test_cpp/../third_party_includes/nanoflann.hpp:1823:16: error: âlog2â is not a member of âstdâ
treeCount = std::log2(maximumPointCount);

the nanoflann.hpp treeCount = std::log2(maximumPointCount); is incompatible with something?

Hello,

Thanks for the code. I had some issues during installation, and want to put it here in case others have similar issues.

I tried to install on Ubuntu 18 using conda, and had to make the following changes to setup.py:

1. add import numpy
2. add numpy.get_include() to include_dirs

Thanks.

when i compile the normals_hough,it comes
E:\normals_Hough-master\normals_Hough-master\include\Normals.h(252) : error C2664: “nanoflann::KDTreeSingleIndexAdaptor<Distance,DatasetAdaptor,DIM,IndexType>::knnSearch”: 不能将参数 3 从“long *”转换为“__w64 int *”
how can i solve this problem

Hello,

Thanks for sharing the source code. I am now comparing different normal estimation methods.
But would you mind to refine the README for running example of "Fast and Robust Normal Estimation"? There are several errors when compiling.

Best,
Yumin

Dear Alexandre,

There seems to be an inconsistency in the methods of cubes_radius() at L1682 and Unif_radius() at L1418 where the radius2 is used as it is from the input argument, while in two other methods, e.g. cubes_knn() and the Unif_knn() it is computed as follows (L1287 or L1573): radius = sqrt(radius);
Is this a bug?

Best regards
Dmytro Bobkov