Thx for this database. Odometry is a lot better now. One problem I see when opening it on my side is that g2o (which is the optimizer used) breaks the map at some point. All poses are 2D and the graph is optimized in 2D, but we get errors in Z like this: #1169
Thx for this database. Odometry is a lot better now. One problem I see when opening it on my side is that g2o (which is the optimizer used) breaks the map at some point. All poses are 2D and the graph is optimized in 2D, but we get errors in Z like this:
When selecting GTSAM instead we have the correct optimization in 2D:
However, the graph optimized with g2o in databaseViewer doesn't suffer the same problem than the one optimized by rtabmap standalone. I'll have to look in more depth to find the difference.
For reference, here is the map optimized with GTSAM including landmarks:
One small detail is the poor depth estimation on the floor by the zed camera, which creates points under the ground, though it is more a sensor issue (or with disparity approach used by zed):
When selecting GTSAM instead we have the correct optimization in 2D:
However, the graph optimized with g2o in databaseViewer doesn't suffer the same problem than the one optimized by rtabmap standalone. I'll have to look in more depth to find the difference.
For reference, here is the map optimized with GTSAM including landmarks:
One small detail is the poor depth estimation on the floor by the zed camera, which creates points under the ground, though it is more a sensor issue (or with disparity approach used by zed):
cheers, Mathieu
Originally posted by @matlabbe in https://github.com/introlab/rtabmap/issues/384#issuecomment-486460380