search

lmb-freiburg / freihand

A dataset for estimation of hand pose and shape from single color images.
https://lmb.informatik.uni-freiburg.de/projects/freihand/
379 stars 77 forks source link

root recover #8

Open moranli19 opened 4 years ago

moranli19 commented 4 years ago

How to generate xyz_root from uv_root on evaluation set?

I have noticed that the xyz_root can be recovered from uv_root using the functions in your code (https://github.com/lmb-freiburg/freihand/blob/master/utils/model.py#L58) with 'scale' parameters which are parts of mano parameters[60]. It's obvious that such 'scale' is different from the bone length of each sample.

I want to know the exact meaning of 'scale' (mano parameter[60]). And the way to calculate xyz_root from uv_root on the evaluation set ('scale' parameter is not provided).

It seems that the root joint precision influences the final results (kp_mean_error (aligned)) on Codalab when using the Latent25D method[1]. So, it is important to generate a precise root joint for each evaluation sample. Or, the alignment on Codalab should remove the influence of xyz_root.

Thanks for any reply!!!

[1]: Iqbal et.al. Hand Pose Estimation via Latent 2.5D Heatmap Regression.

samiroor commented 4 years ago

Hi @moranli19 , did you get your answer?

www516717402 commented 4 years ago

@sameeroor @moranli19 ManoScale means proportion between Z{id=9,camera space} and focal length from Mano files. However another scale indicate length of idex=[9,10] from scale files. About Mano scale, you need see source code to comprehensive.https://github.com/lmb-freiburg/freihand/blob/0808a4d29107ebcca2724b135728c1aa72022b77/utils/model.py#L58