Generate parsed data
pangyyyyy opened this issue · comments
Hi, thanks for the great work! I'd like to check how did you generate the parsed data from the original datasets? And if you mind providing the codes for them?
Thank you!!
which datasets do you want?
Hello! Thanks for your reply
I'm keen to understand how you generate the parsed data esp for 3DPW and MuCo datasets. Would you be willing to share the preprocessing steps/scripts for these datasets?
I'm afraid I cannot share the codes, but it is not very difficult to generate such json files.
You can open the json files and understand the dictionary structure of them.
No worries! I tried to derive the same values as your 3DPW parsed data but I was unable to obtain the same values for fitted_joint_cam. You mentioned in your readme that the ‘fitted_joint_cam is 24x3 joint coordinates (SMPL joint set) in camera-centered 3D space (meter unit), which is regressed from the SMPL mesh.
I have a few questions:
(1) What are your inputs into the SMPL regressor?
(2) Did you use a gendered or gender-neutral regressor?
(3) Was any translation applied?
(4) Did you offset the SMPL outputs by root coordinates obtained from RootNet?
Hope to get your advice on this, thanks!
I forward SMPL parameters, provided in original 3DPW dataset.
However, translation vectors in the original 3DPW are incorrect.
Therefore, I performed rigid alignment the output of SMPL parameters to 'jointPositions', provided in 3DPW dataset.
I used a neutral gender.
Thanks for the reply! Do you mind elaborating on how you performed "rigid alignment the output of SMPL parameters to 'jointPositions'"?
I forwarded the SMPL betas and poses and used a gender neutral model. I tried to align my SMPL outputs using the root_coordinate of the jointPositions, provided in 3DPW dataset, but I still got very different values from fitted_joint_cam. I looked at the root joint coordinates obtained from RootNet and aligning the SMPL outputs to the root coordinate provided seems to give similar values to those in fitted_joint_cam.
For instance:
For annotation id: 32664
Actual offset needed to be applied to obtain `fitted_joint_cam` values from derived SMPL outputs: [1.1488, 0.2267, 5.6820]
Root coordinate of `jointPositions`: [2.9659, -0.3631, -5.8923]
Root cam coordinate obtained from RootNet: [1.1495038270950317, 0.28018078207969666, 5.689423084259033]
You should also consider camera extrinsics.
Here is the rigid alignment codes, which get 3D rotational matrix (include rotation/scale/translation).
I performed rigid alignment of SMPL outputs to jointPositions but was still unable to get fitted_joint_cam values
Sorry for the trouble, can I check if this is the correct procedure?
For annotation id: 32664
keypoints_3d = data['jointPositions'][0][0].reshape(24, 3) # 3D keypoints
smpl_joints = np.array(output.joints).reshape(24, 3) # SMPL outputs
aligned_joints = rigid_align(smpl_joints, keypoints_3d) # align SMPL outputs to 3D keypoints using provided code
aligned_joints:
array([[ 2.96633185, -0.35136265, -5.88738065],
[ 3.0315059 , -0.44573565, -5.85921113],
[ 2.90759431, -0.44275249, -5.92828365],
[ 2.97852218, -0.2397811 , -5.90929941],
.....
fitted_joint_cam:
array([[ 1.14881543, 0.2267409 , 5.68205287],
[ 1.21106176, 0.3300526 , 5.6515299 ],
[ 1.09256051, 0.31742936, 5.74177306],
[ 1.16648478, 0.10804464, 5.69278809],
....
smpl_joints:
array([[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00],
[ 6.22463301e-02, 1.03311732e-01, -3.05229668e-02],
[-5.62549196e-02, 9.06884819e-02, 5.97201772e-02],
[ 1.76693555e-02, -1.18696287e-01, 1.07352212e-02],
....
keypoints_3d:
array([[ 2.96588857, -0.36313492, -5.89231832],
[ 3.02422186, -0.4446724 , -5.8830483 ],
[ 2.92139617, -0.45167833, -5.93622714],
[ 2.98745489, -0.23942477, -5.91659344],
.....
fitted_joint_cam - fitted_joint_cam[0] = smpl_joints
I initially thought that the problem might lie in finding the correct root coordinates. I was wondering if the root coordinates from RootNet were used?
The root coordinates from RootNet were predictions, so not proper to use them as GTs (json files include GTs).
jointPositions are in the world coordinate system, so you should apply camera extrinsics ('cam_poses') to make them in the camera-centered coordiante system.
Please try projecting coordiantes into image space and visualize them for the debugging.
OK thanks for the clarification!
I obtained R, t from camera extrinsics and transformed jointPositions to cam coordinate system. Applying rigid alignment to the SMPL outputs and cam-aligned keypoints give an output close to (but not exactly the same as) the "g-t" fitted_joint_cam.
array([[ 1.14638129, 0.21608887, 5.68087656],
[ 1.20532955, 0.31419452, 5.65175911],
[ 1.09309508, 0.30212931, 5.73773498],
[ 1.16324552, 0.10342685, 5.69099316],
....
I will try to visualise them for further debugging, thanks!
Sure :)