search

RonaldSun / VI-Stereo-DSO

Direct sparse odometry combined with stereo cameras and IMU
393 stars 124 forks source link

You have mixed different SE3 Jacobian types #30

Open tanjunyao7 opened 4 years ago

tanjunyao7 commented 4 years ago

DSO evaluates the Jacobian of the photometric error with left Jacobian of SE3. The update step of poses is also done in the left-multiplication way. However, your jacobians of IMU residuals, which probably originates from https://github.com/jingpang/LearnVIORB, adopt right Jacobian. I just wonder why the optimization still converges.

studennis911988 commented 4 years ago

The author derived the right to left Jacobian, which is also mentioned in issue28. However, I have no idea how he derived that.

tanjunyao7 commented 4 years ago

The author derived the right to left Jacobian, which is also mentioned in issue28. However, I have no idea how he derived that.

OK, I see. I havn't read through the code before. The deriviation may be formulated as follows: _exp(d_xi_l) T = T exp(d_xir) _exp(d_xi_l) = T exp(d_xi_r) T^(-1) = exp(Adj(T) d_xir) _d_xi_l = Adj(T) d_xir

studennis911988 commented 4 years ago

@tanjunyao7 Your derivation really makes sense, thanks a lot!

DengueTim commented 2 months ago

But optimisation of the transform from the IMU World frame to DSO's frame uses the right Jacobian right? I'm struggling to understand the intuition of eq. A.6 in Stumberg's paper's supplement...