PCL中的特徵區別 3DSC和SHOT，
阿新 • • 發佈：2019-01-29
the USC descriptor is basically the 3D Shape Context descriptor (as
described in Frome et al. , "Recognizing Objects in Range Data Using
Regional Point Descriptors",
where a different definition for the local Reference Frame of each
feature being described is used. As explained in the original 3DSC
paper, due to the ambiguity in orienting their local RF, they need L
different rotations for each point, leading to L different descriptors
(L=12 in their experiments). In USC, the same description method is used
but relying on a local RF which is unique, thus it leads only to one
description for each point.
Then, your question is probably what is the difference between the
description method of 3DSC/USC and that of SHOT (without texture). The
common aspect is that they both place an oriented 3D grid around the
feature point. Then, 3DSC/USC accumulates the number of points falling
in each spherical sector based on their 3D coordinates: hence, it is
basically a 3D histogram. Each inserted point doesn't count 1 in the
histogram, but it is weighted based on the local point density around it
and on the volume of the spherical sector it falls in. Instead, SHOT
computes an histogram for each grid sector: this histogram accumulates
the angle between the feature point and all points falling in that
particular sector. Quadrilinear interpolation is then used to smooth out
the sampling effect of the histogram. Then, all computed histograms (one
for each sector) are juxtaposed to form the final descriptor vector.
described in Frome et al. , "Recognizing Objects in Range Data Using
Regional Point Descriptors",
where a different definition for the local Reference Frame of each
feature being described is used. As explained in the original 3DSC
paper, due to the ambiguity in orienting their local RF, they need L
different rotations for each point, leading to L different descriptors
(L=12 in their experiments). In USC, the same description method is used
but relying on a local RF which is unique, thus it leads only to one
description for each point.
Then, your question is probably what is the difference between the
description method of 3DSC/USC and that of SHOT (without texture). The
common aspect is that they both place an oriented 3D grid around the
feature point. Then, 3DSC/USC accumulates the number of points falling
in each spherical sector based on their 3D coordinates: hence, it is
basically a 3D histogram. Each inserted point doesn't count 1 in the
histogram, but it is weighted based on the local point density around it
and on the volume of the spherical sector it falls in. Instead, SHOT
computes an histogram for each grid sector: this histogram accumulates
the angle between the feature point and all points falling in that
particular sector. Quadrilinear interpolation is then used to smooth out
the sampling effect of the histogram. Then, all computed histograms (one
for each sector) are juxtaposed to form the final descriptor vector.