search

thodan / bop_toolkit

A Python toolkit of the BOP benchmark for 6D object pose estimation.
http://bop.felk.cvut.cz
MIT License
376 stars 135 forks source link

Questions about megapose dataset #100

Closed parkjaewoo0611 closed 9 months ago

parkjaewoo0611 commented 9 months ago

Hi, Thanks for releasing great dataset,

I am setting up a dataset following the guide provided at https://github.com/thodan/bop_toolkit/blob/master/docs/bop_challenge_2023_training_datasets.md, and I have two questions.

According to the guide, the .obj files downloaded from https://research.google/resources/datasets/scanned-objects-google-research/ are not centered at the origin. When I projected them in this state, they did not match the images. However, when I adjusted them to make the centroid the origin, they seemed to align better. Is this adjustment necessary?

When I adjust the centroid to the origin, the direction of the translation vector to the object seems correct, but the scale is off. When I scale the mesh by 0.1 as instructed in the guide, it appears too small. Are there any additional scale considerations I should be aware of? (I've already tried scaling from mm to m and vice versa, but the difference in scale seems more subtle than such large discrepancies, similar to the difference shown in the attached image.)

image

parkjaewoo0611 commented 9 months ago
  1. centroid -> origin
  2. scaling max(x, y, z) to 0.1 resolves the issue. Since the problem has been resolved, close this issue.
ylabbe commented 9 months ago

Thanks a lot for pointing out that our instructions to scale the mesh are incorrect. See also #98.