Interpreting 2d coordinate output
riboyuan99 opened this issue · 8 comments
I am able to get 2d coordinate by following code:
poses_3d, poses_2d = parse_poses(inference_result, input_scale, stride, fx, is_video)
and the output looks like this:
[ 9.87000000e+02 8.40000000e+01 7.44423568e-01 -8.00000000e+00
-8.00000000e+00 -1.00000000e+00 -1.00000000e+00 -1.00000000e+00
-1.00000000e+00 9.95000000e+02 8.40000000e+01 7.20652282e-01
1.02900000e+03 1.43000000e+02 4.71996784e-01 1.02900000e+03
1.85000000e+02 4.42449659e-01 9.95000000e+02 2.10000000e+02
6.77076042e-01 9.61000000e+02 2.86000000e+02 6.98154628e-01
9.53000000e+02 3.54000000e+02 5.87647200e-01 9.53000000e+02
8.40000000e+01 6.99335754e-01 9.28000000e+02 1.43000000e+02
3.37824076e-01 9.28000000e+02 1.51000000e+02 1.61912560e-01
9.61000000e+02 2.10000000e+02 6.83217466e-01 9.87000000e+02
2.86000000e+02 6.84145570e-01 9.95000000e+02 3.45000000e+02
5.94412923e-01 -8.00000000e+00 -8.00000000e+00 -1.00000000e+00
-8.00000000e+00 -8.00000000e+00 -1.00000000e+00 -8.00000000e+00
-8.00000000e+00 -1.00000000e+00 -8.00000000e+00 -8.00000000e+00
-1.00000000e+00 1.42583218e+01]
How do I know which body part it is referring to?
Any advice would be appreciated.
Hi! You can check draw_poses, which parses and draws 2D poses coordinates.
Thanks for your helpful hint. I have read the code but got stuck on one thing.
We only have 18 body parts:
But we have 19 kpt_id:
for kpt_id in range(pose.shape[1])
What is the missing one?
Any hint would be appreciated
Nothing is missed. 2D network detects 18 keypoints you have found above. The order of 2D keypoints is aligned to match the 3D network output. 2 is reserved value for pelvis in 3D network. Now it is not predicted, it was left for the future.
Thank you very much!
@Daniil-Osokin sorry so there is no predicted 3D pelvis in this model?
Right.
@Daniil-Osokin but parse poses emits a pelvis in index 2? Is that not an accurate calculation or something?
It always has "-1" as coordinates values. It is left for the legacy reasons.