Dennis Cychuang

Yes, you are right, since the visual reference frame has z axis pointing forward, so you need some modification. `` Eigen::Quaterniond quat(frame->camToWorld.rotationMatrix()); `` `` q_.coeffs()[0] = q_.coeffs()[0]; // w->w ``...

There is truly some bugs in current code, especially in initializer part.

Since DSO applies FEJ to the system, the marginalized Hessian (HM) is kept at the same linearlization point, but we could still use first order Tayler expansion to keep updating...