Ensure good accuracy RGBD odometry

[pedrotop3d]

Hi All,
I´m trying to obtain the better accuracy possible in the odometry using primesense. Using the default parameters, i can see many descriptors, even when the texture is very homogeneous; i suppose a lot of those descriptors cannot generate good quality matching.

What i´m looking for is use only the most reliable descriptors points (which generate only good matching) and refuse the bad ones (or the doubth cases). I´m aware, increasing the quality will reduce the choosed points and it will loose the odometry soon, but i can assume this.

Can someone suggest which parameters can i try to increase ensure a good matching and a good odometry accuracy?

Thanks in advance!
 

 

[matlabbe]

Hi,

Most descriptors work fine for odometry, it is more textureless environments that is causing most tracking lost problems. Increasing GFTT quality parameter may give less features but more discriminative ones, so better matches. Lowering NNDR of visual registration may give better matches too.

cheers