search

chungyiweng / humannerf

HumanNeRF turns a monocular video of moving people into a 360 free-viewpoint video.
MIT License
786 stars 86 forks source link

coordinate in wild data #18

Closed caiyongqi closed 2 years ago

caiyongqi commented 2 years ago

Hi, it's a outstanding work. I'm a little confused about some details in the code.

  1. Why the origin of the SMPL coordinates in the wild data is at 0-joint, which doesn't seem to be the case in the zju-mocap data, and what effect does this difference have? https://github.com/chungyiweng/humannerf/blob/7389c4049c0d5b14afdbac41a605ddad9d51dd4a/tools/prepare_wild/prepare_dataset.py#L73 In the zju-mocap data, Rh and Th are the transformation of the smpl coordinate system (without global R, that is, pose[0]=[0,0,0]) to the world coordinate system. However, in the wild data they are set to pose[0, 0:3] and joints[0] respectively, https://github.com/chungyiweng/humannerf/blob/7389c4049c0d5b14afdbac41a605ddad9d51dd4a/tools/prepare_wild/prepare_dataset.py#L69
  2. which seems to mean that the new SMPL coordinate system origin is at joints[0] position, the world coordinate system (obtained by Rh and Th) is the old SMPL coordinate with global R.
chungyiweng commented 2 years ago

Hello,

Your understanding is correct. I did this just for ease of implementation -- by doing so, I can directly rotate the human model (or equivalently, rotate the camera) without the need of pre-shifting the model to the center of rotation (i.e., the joints[0]).