fast hog资料整理:
https://github.com/joaofaro/FHOG
https://github.com/ppaanngggg/fhog
这个有两种:
1.DPM的
Mat tmp_image;
image.convertTo(tmp_image, CV_32FC1);
cout<<image.rows<<" "<<image.cols<<endl;
Mat fhog_feature = fhog::fhog(tmp_image);
cout<<fhog_feature.rows<<" "<<fhog_feature.cols<<endl;
这个很直接,直接计算返回一个矩阵
2.from pdollar's toolbox,这个在pc上使用,因为有sse指令加速,在arm上也可以使用,有neon加速。
参数说明:
Inputs: float* I - a gray or color image matrix with shape = channel x width x height 图像数据 int *h, *w, *d - return the size of the returned hog features 返回的特征尺寸 int binSize -[8] spatial bin size cells大小 int nOrients -[9] number of orientation bins 方向 float clip -[.2] value at which to clip histogram bins 大于0.2的特征值截断 bool crop -[false] if true crop boundaries 裁剪边缘的特征 Return: float* H - computed hog features with shape: (nOrients*3+5) x (w/binSize) x (h/binSize), if not crop
kcf里面的fhog检测:
https://github.com/TianyeAlex/KCF_vs
int getFeatureMaps(const IplImage* image,constint k, CvLSVMFeatureMapCaskade **map)