search

themightyoarfish / deepVO

:video_camera: Tensorflow implementation of RCCN visual odometry by Wang et al.
50 stars 17 forks source link

Information about the dataset #2

Closed arpitg1304 closed 6 years ago

arpitg1304 commented 6 years ago

Hi,

Can you please tell us about the structure of your dataset like what are the sequences , what is chunk size and all the parameters used to create the batches. It is bit confusing to understand without looking at the dataset. It would be really helpful if you could tell us about the structure of your data

Thanks

themightyoarfish commented 6 years ago

The dataset consisted of 6-d poses and consecutive camera images. We cut the entire sequence into chunks of some length (the sequence_length) and each pose within a subsequence was relative to the initial pose of the subsequence (not the initial pose of the robot, but the data manager will do that for you). We tried different subsequence lengths but I can't give you a good value. We managed any kind of convergence only on a test dataset where the robot moved only forwards and backwards randomly.

491734045 commented 6 years ago

Hi, Have you gotten a good result like the original paper showed using KITTI data for training and testing?

themightyoarfish commented 6 years ago

Unfortunately, we didn't get to using the actual KITTI data since our model didn't even train properly on our toy datasets.