KITTI splits - Odometry splits

[jahaniam]

Hi,

Since it was not mentioned in our github, I thought maybe it is useful to say here. If you are using deep neural network depth estimation in KITTI, just one thing to doublecheck:

KITTI raw dataset has many splits and these splits have overlap with each other. It is important that you don't test on the data which was in your training (or mention it in the paper that which sequences was in the training set).

KITTI Stereo 2015 (know as KITTI split in monodepth paper): 00 yes 01 yes 02 yes 03 no 04 yes 05 yes 06 yes 07 yes 08 yes 09 yes 10 yes

Eigen Split: 00 no 01 yes 02 yes 03 no 04 no 05 no 06 yes 07 no 08 yes 09 yes 10 yes

"yes" means that the odometry sequence is in the training set. You can double check it yourself as well and find the mappings.

There is also another split used by SfMLearner which splits the odometry sequences 00-10 into test and train (I think 00-07 for train and 08-10 for test. see their paper)

What we did was we excluded the odometry sequence 00 -10 from the whole KITTI raw dataset and trained on that. Our pre-trained model will be released in our GitHub soon (hopefully in few days) in case other people want to just use it and evaluate their methods.

Thank you, Ali

[yan99033]

The pre-trained frozen graph is now released. Please go to the monodepth-cpp github page to download the models.

[manymuch]

Hi Jahani and Yan,

Do you know where can I find the official source of the relation between eigen split and odometry? Because I found the statement is inconsistent with D3VO in which they state that

sequences 00, 03, 04, 05, 07 are in the training set of the Eigen split

I found this is important for benchmark comparison in model-based VO. Thank you!

[manymuch]

I believe your are right, I just manually check odometry sequence9, which is 2011_09_30_drive_0033_sync in KittiRaw, and appears in Eigen Split.