JimmySuen / integral-human-pose

Integral Human Pose Regression
MIT License
471 stars 76 forks source link

[ Integral Pose Regression Implementation ] #41

Open IemProg opened 3 years ago

IemProg commented 3 years ago

Hi,

Wonderful work, and thanks for the repo !

I'm quite new to Pose estimation field, I'm having a hard time to understand why did you multiply by a tensor (Torch.arange) in _generate_3d_integral_preds_tensor(heatmaps, num_joints, x_dim, y_dim, zdim) function at link after calculating the heatmaps ?

Furthermore, how can we calculate the variance of the heatmaps ?

Thanks !

lck1201 commented 3 years ago

Torch.arange refers to the heatmap location, i.e., [0, 63]. So accu_x is a weighted sum of possibility and location.