Firstly I would like to thank you for such a powerful SLAM solution. I have two questions regarding the flag RGBD/LocalBundleOnLoopClosure.
1. I enabled the flag and executed the clips under the ‘Robot’ category of TUM dataset in the ‘RgbdDataset’ tool to test the accuracy of the obtained trajectory. I noticed a degradation with respect to results with disabled bundle adjustment on loop closure. Please advise.
2. In case we enabling the RGBD/LocalBundleOnLoopClosure flag, will it be helpful to also optimize the poses of all the nodes participating in the bundle adjustment, rather than optimizing just the pose of ‘toS’ node only (While considering that the poses were transformed relatively to ‘fromS’)?
Thanks in advance,
1. `RGBD/LocalBundleOnLoopClosure` will use neighbor nodes sharing same visual features to do a local bundle adjustment, which would improve loop closure accuracy, at cost of using more computation time.
2. Unless odometry is less accurate (locally) than loop closure transformation estimation, I would expect that local bundle adjustment would add more errors to odomery links than remove them. However, I guess we could experiment or add option to update all poses contributing to local bundle adjustment, but I kinda like keeping the original constraints in the graph for debugging. And if odometry is indeed more accurate locally than loop closure constraints, it is preferable not touching them.