search

ClementPinard / SfmLearner-Pytorch

Pytorch version of SfmLearner from Tinghui Zhou et al.
MIT License
1.01k stars 224 forks source link

about the pose scale #154

Closed everythoughthelps closed 10 months ago

everythoughthelps commented 10 months ago

I run sfmlearner on tanks dataset, the visualization of depth estimation looks reasonable. gt: image pred: image However, the pose doesn't look good: pred: 120 poses, 0.241m path length gt: 120 poses, 13.809m path length

ClementPinard commented 10 months ago

The scale is not guaranteed. That's why the script to test pose is rescaling the path.

If you really want want to have absolute path length, you need a real world anchor. Either the depth (multiply the path length by the ratio between predicted depth and groundtruth depth) or the path itself, and the evaluation will just be about relative error

There is intrinsically a ambiguity regarding scale factor because you cannot know if you are looking at a small scale replica with small camera displacement or a full scale one with regular camera displacement, everything looks exactly the same.

everythoughthelps commented 10 months ago

Thank you for your response!