Datasets:
v1.0/v1.1 BOP pose convention — apply diag(1,-1,-1) when scoring against BOP toolkit
#1
by zeredata-admin - opened
Heads up for anyone using this dataset against the BOP toolkit or other BOP-standard pose scoring:
The v1.0 and v1.1 releases export object pose extrinsics (cam_R_m2c, cam_t_m2c in scene_gt.json) in OpenGL convention (negative-Z forward) rather than the BOP-standard OpenCV convention (positive-Z forward).
To score against BOP toolkit baselines, apply a diag(1, -1, -1) transform to the rotation matrix when consuming the dataset:
import numpy as np
R_cv = R_gl @ np.diag([1, -1, -1])
The producer-side fix is in progress; v1.2 will ship with BOP-canonical extrinsics out of the box. The methodology has been disclosed in our internal benchmark results as well.
If you've hit this in your own work or have questions, happy to discuss here.
— Umit