search

gogojjh / M-LOAM

Robust Odometry and Mapping for Multi-LiDAR Systems with Online Extrinsic Calibration
http://gofile.me/4jm56/zU2yvg3bH
473 stars 88 forks source link

about the edge residual #4

Closed nonlinear1 closed 3 years ago

nonlinear1 commented 3 years ago

in A. can-Based Motion Estimation of VI.INITIALIZATION in your paper:"Robust Odometry and Mapping for Multi-LiDARSystems with Online Extrinsic Calibration". I was confused with the edge residual. for "where[w1,d1]and[w2,d2]are the coefficients ofΠ1andΠ2,w1coincides with the projection direction fromLtop, andΠ2is perpendicular toΠ1s.t.w2⊥w1, andw2⊥L. " in your paper, I think p in Π2, so rH(x,p,Π2) is equal to zero for ever! Could you give me some explantaiton about this part? Thank you very much!

gogojjh commented 3 years ago

Thanks for your question.

The pipeline of the motion estimation should be: step1_each edge feature finds its correspondences -> step2_iteratively optimize the residual function with several iterations. At step 1, rH(x, p, Π2) is zero; but during the optimization, since x is changing, rH(x, p, Π2) is not always zero. This term can constrain that the optimization direction goes into the right direction.

But we find that in many experiments, the usage of rH(x, p, Π2) only has a little help to the precision, but induces a higher computation cost. Therefore, the currently released code does not use rH(x, p, Π2) anymore. The new residuals are detailed in: https://ram-lab.com/file/site/m-loam/m-loam-gf/supp.pdf

nonlinear1 commented 3 years ago

Thanks a lot! I understand totally!