search

ethnhe / PVN3D

Code for "PVN3D: A Deep Point-wise 3D Keypoints Hough Voting Network for 6DoF Pose Estimation", CVPR 2020
MIT License
489 stars 105 forks source link

gt pose of meta file in YCB dataset #94

Closed hetolin closed 2 years ago

hetolin commented 2 years ago

Hi, thanks for sharing your code! The pose array from the metafile is strange, they have a shape with(3,4), (4,3) and (4,4).

For example, in the scene 00001/000005-meta.mat, the meta['pose'] = [[ 0.994375 -0.101504 0.0302549 0.07000585] [-0.0357059 -0.590176 -0.806485 0.13960156] [ 0.0997172 0.800868 -0.59048 1.08210617]] [[ 0.99682217 -0.66709934 -0.98591953 0.67320229] [-0.07169833 -0.74246254 0.1671971 0.73916946] [ 0.03472134 0.06106711 -0.00285112 0.02068367] [-0.0321408 -0.07229631 0.1321786 0.17324216]].

However, the ycb_dataset.py read the pose following the way below: r = meta['poses'][:, :, i][:, 0:3] t = np.array(meta['poses'][:, :, i][:, 3:4].flatten()[:, None]) As for the shape with (4,4) then the t-vector=[0.67320229, 0.73916946, 0.02068367], it so strange as the value of z-axis is too small.

ethnhe commented 2 years ago

The meta info is created by the authors of PoseCNN, who create the YCB dataset. I think the pose data is in shape (3, 4, n_instance) and our processing code is correct.

hetolin commented 2 years ago

thanks for you prompt reply, I will check it