Pose and pose_inv differ significantly about sc-sfmlearner-release HOT 3 CLOSED

Hi,

Thanks for your question. For reproducing, I do it in test_vo.py (kitti sequence 09 using resnet18_depth_256 model.). There is no data augmentation as in training. so it would be more clear. I debug by adding following codes, and results show that the t3-error between forward and backward is about 2e-3. Does it match your finding?

from sc-sfmlearner-release.

Yes, that matches what I have. There is a consistent difference in translation of about 10-25% for all samples, which seems like an issue. I was wondering if you had any insight as to why this would occur - possibly there is some mechanism that causes the loss function to increase if more 'black' space exists on the reconstructed images?

from sc-sfmlearner-release.

It should be consistent, but it is very difficult for the network to predict very consistent results. If you find this issue important, you may minimize this error during training as a loss function. I did this before, and the loss looks small and impact little for the training.

from sc-sfmlearner-release.