About Pose Results about sfmlearner-pytorch

[ClementPinard]

Hello,

yes you are right ! the data separation between train, validation and test is supposed to follow Eigen's split, which means that, as you said, sequences 9 and 10 are either in training or in valid.

Unfortunately, not much more is done to avoid overfitting on these scenes, as the split between train and test is random (whether it is from this repo or Zhou's repo). It could be nice to explicitely dismiss sequences 9 and 10 from being in the training set.

It appears that sequences 9 and 10 are indeed considered test sequences in the odom layer here, but I am not sure what was the network used when testing pose, was the network trained on Odom (0-8) when testing pose and trained on Eigen split when testing depth ?

The pose results on README where obtained with a training on a random split so it might have overfitting.

I am aware that this test is not very thorough, because my personal goal was to obtain good results on depth.

[cashiwamochi]

Thanks!
As you said, I got good results on depth.
I think that, the network trained on Odom(00-08) was used in testing pose and the network trained on Eigen split was used in testing depth, but I'm not sure, too...
I will try what I wrote.