Closed biggiantpigeon closed 10 months ago
You are asking why we don't directly initialize the sliding window and features in addition to the inertial 15dof state. This is mainly because recovering the initial covariance of the whole system is expensive, so this will likely be a bottleneck. I think I have some TODOs in the code as it would be nice to recover the features and directly use them. The system should be more stable in this case, but not sure.
Oh I see, thanks a lot~
I notice that poses and features optimized from dynamic initialize not used, as in code
this means all poses from the time before init success is cleared in sliding window(and also the features in database), and we rebuild the sliding windows afterwards. Why don't we use these poses and features? Any concern? As I recall, both orbslam3 or VINS-mono did not discard them. Thanks!