narf descriptor

asked 2011-03-05 04:40:45 -0500

Yianni
123 ●6 ●10 ●14

Hi,

I was looking at the source code of the NARF descriptor and I modified the tutorial example (narf_feature_extraction) to calculate all the time the NARF descriptors from a pointcloud coming from a kinect camera. I have noticed a couple of things:

Some times there are more descriptors than keypoints... I see that the size of the keypoints is taken from keypoint_indices2 and the descriptors are held in narf_descriptors. Is it correct that there are more descriptors than keypoints computed?
I noticed the descritptor values for a pointcloud of a mug for example are around between -0.3 and 0.3, and with any object or the complete pointcould the values are between -0.5 and 0.5... If someone has experience and know the range of values it would be great... as I had a look on the paper (Steder et al., 2011) but I couldn't find out something specifically about it...

Many Thanks,
Yianni

edit retag flag offensive close merge delete

add a comment

6

answered 2011-03-06 19:49:12 -0500

Bastian Steder
76 ●2 ●2 ●2

Hi Yianni,

if the rotational invariance of the features is used there can be more than one descriptor per keypoint, if there was more than one strong rotation around the normal found for the extracted range value patch. Additionally, the number of descriptors can be lower than the number of keypoints. E.g., if the resolution in the area was too low.
The values of the descriptor are normalized to -0.5 and 0.5, so that the difference between descriptor cells is between 0 and 1. I typically use the Manhattan distance divided by the descriptor size as the descriptor distance, which will always be in [0,1]. I prefer my values to be in areas where an interpretation is easy, which is why I use this normalization.

Cheers, Bastian