ClementPinard
commented
5 years ago Amen to that :)
GPS rtk is still sometimes very noisy, and its main interest is for long range odemetry evaluation. A good hint about that is the official kitti odometry benchmark which dropped evaluation on sequences that are too short :
http://www.cvlibs.net/datasets/kitti/eval_odometry.php
On 03.10.2013 we have changed the evaluated sequence lengths from (5,10,50,100,...,400) to (100,200,...,800) due to the fact that the GPS/OXTS ground truth error for very small sub-sequences was large and hence biased the evaluation results. Now the averages below take into account longer sequences and provide a better indication of the true performance.
However here are some points that might help you :
- Raw GPS coordinates are noisy, but poses from kitti odometry are somewhat cleaned and have a better quality. You can try to train on kitti odometry training sequences, but that's a smaller dataset.
- Poses can be deduced from GPS but also from estimated speed! It's probably better to integrate the speed to have the displacement since the sequence is very short. There is a snippet available on the kitti testkit to kinda do that :
an-kumar
sconlyshootery
I tried modifying the code to take the ground truth pose as input and use that for the warp rather than learning a PoseNet. I expected this to help convergence since the depth network would not be getting trained on bad poses in early stages. However, it ended up being considerably worse than using PoseExpNet.
I'm still looking into it to make sure I didn't do something wrong, but I was wondering if you ever tried this or have any thoughts. Perhaps the "ground-truth" poses are not that accurate, so learning the pose function ends up being more precise? If that's the case, than the odometry benchmarks are somewhat meaningless..