test_disp.py fails with the recent commits about sfmlearner-pytorch HOT 4 CLOSED

Yes, sorry about that, should have been more careful when pushing.

This has been solved in commit #69588b0 (3 hours ago)

I'm still struggling to get comparable results with original paper so some other change might occur in the next days. Thanks for your vigilance anyway !

from sfmlearner-pytorch.

Thanks. Do these changes in depth evaluation result in different scores too? How similar/different is it to Eigen et al, or Zhou et al?

from sfmlearner-pytorch.

The difference here is that in addition to determining depth scale_factor from ratio of medians, we try to get it from ratio of translations. The main idea is that getting median of GT depth is impossible in real condition, while getting GT displacement is.

Obviously this leads to worse results but it's closer to a real usecase.

Anyway, we compute both from depth ratio and from translation ratio, and when you don't give a pose network, it will only compute with original scale factor.
Besides, with the option --output-dir you can save your inference values in a npy file that you can feed to e.g. Zhou's test_disp script. To be absolutely sure that tests are consistent.

from sfmlearner-pytorch.

Ok cool. Thanks.

from sfmlearner-pytorch.