ylabbe / cosypose

Code for "CosyPose: Consistent multi-view multi-object 6D pose estimation", ECCV 2020.
MIT License
301 stars 89 forks source link

Converging on identity matrix #71

Closed murrdpirate closed 1 year ago

murrdpirate commented 2 years ago

I'm working on implementing CosyPose on my own data of a single object. I had to modify the code a bit to deal with my own formatting, but it's close to the same.

While the network does learn, it seems to very quickly learn to output the identity matrix for the pose refinement, but does not improve beyond that. I suppose the identity matrix is a big local minimum, as it is much better than a random refinement.

Is this something that was seen with CosyPose? Any trick to avoid it?

murrdpirate commented 2 years ago

It appears that pretrained weights (from imagenet) reduces my final error by about half. It doesn't appear that cosypose uses pretrained weights by default. There may be other important factors as well.

murrdpirate commented 1 year ago

It looks like K_crop was computed incorrectly. Fixing that seems to have resolved my issue.