Improvement Of The Curvature Computation
(VBclav HlavBE, TomBB Pajdla; Milos Sommer)
The curvature computed at every outline point can be used as a local description of 2D curved objects. The method for obtaining the curvature from the noisy outline and for dividing the outline into segments with an equal change of curvature was presented before. We report the improved solution to the curvature computation problem here. We have reimplemented Lowe?s method. We have applied performance analysis and the discrete nature of the data has been considered. Based on the analysis we have achieved considerable improvements of the method that eliminates significant part of systematic bias. We show which part of the bias is due to smoothing and which is caused by other phenomena like the anisotropy of the raster, lim- ited size of the Gaussian, numerical integration of the convolution, and discretization. The improvement of computing the curvature of the digitized curves is presented in this paper.The standard scheme, i.e. computing curvature using convolution with the truncated Gaussian kernel, was studied. First, they show that that systematic bias caused by curvature can be removed. Second, they demonstrate that large portion of the error has roots in other phenomena (i.e. anisotropy of the convolution, Gaussian, numerical integration and discretization).