Closed sunmengnan closed 2 years ago
We have this page to collect model zoo requests. Could you please add your request there following the template? Thank you.
Are these 2d-3d Videopose3d models trained by yourself? Or downloaded
All models in the model zoo are trained using MMPose (unless otherwise specified).
Thx, are these video3dpose models trained by supervised setting, or semi-supervised setting?
We provided both. Please refer to videopose3d_h36m.md for details.
One thing makes me confused is that why only use 10% of S1 for supervised traning in semi-supervised setting? As semi-supervised training use S5,6,7,8., can we use like 90% of S1?
We follow the experiment settings in the paper. You could conduct experiments with your own settings by modifying the config file.
Have you tested these models on some other indoor videos? when I used these models tested on myown videos, when performing the same action(such as walking), the depth joints are not stable, sometimes people stand nearer the camera, the more deeper between head and knees, feet. when people are far from camera(like more than 5 m), the depth among all joints are close to each other.
This could be caused by the gap between training data and the testing scenario. The training data (e.g. human3.6m) is usually collected in constraint settings (environment, human identities, action, camera parameters ...). So the model may not be able to perform well under user settings.
It helps to improve the accuracy and stability if you could normalize your data (e.g. bbox scale) to a similar distribution of the training data.
could you provide model in Causal mode in Videopose3d ?