Open mayank64ce opened 1 year ago
I meant the output of the original OpenPose is a json file with 2d keypoints which look something like this:
{
"version": 1.3,
"people": [
{
"person_id": [
-1
],
"pose_keypoints_2d": [
535.734,
246.827,
0.855832,
....
],
"face_keypoints_2d": [
496.62,
230.471,
0.699601,
...
],
"hand_left_keypoints_2d": [
751.966,
512.104,
0.542437,
...
],
"hand_right_keypoints_2d": [
266.106,
510.039,
0.540713,
....
],
"pose_keypoints_3d": [],
"face_keypoints_3d": [],
"hand_left_keypoints_3d": [],
"hand_right_keypoints_3d": []
}
]
}
Can I get these values, not just the output plot from controlnet_aux ?
This functionality is in the A1111 implementation, e.g. see https://github.com/Mikubill/sd-webui-controlnet/blob/main/annotator/openpose/__init__.py#L122
So if I pass the poses
from https://github.com/patrickvonplaten/controlnet_aux/blob/master/src/controlnet_aux/open_pose/__init__.py#L220
to that function, I should be able to get the json right ?
UPDATE: I tried it and it kinda works. However, the original OpenPose gives 75 values under the pose_keypoints_2d
key whereas controlnet_aux is giving 54 values. Is it detecting less points ?
UPDATE 2: Seems like I need BODY_25 but I am not sure how to use it with controlnet.
There's a good chance that the openpose implementations have diverged. I last updated all of them on 5/27 so it's worth a pass to review.
yeah latest openpose has 7 additinal joints. The pelvis + 2 toes and ankle for both feet. Anyways those didn't matter to me. I approximated the position of pelvis with hip joints and mapped the toes and ankle to feet.
Sorry I don't understand the question