CARLA Skeleton

Pedestrian skeletons in CARLA consist of a root point and 25 joints. Below is an image with the annotated keypoints (CARLA skeleton in T-pose).

CARLA T-pose

The pose can be saved in different coordinate systems:

  1. Using the world coordinates for each joint. This is the ‘world‘ pose.
  2. Relative to the actor’s pivot (usually the center of the bounding box). This is the ‘component‘ pose.
  3. Each joint’s position is relative to the preceding one in the kinematic tree (so the crl_hips__C position is described relative to crl_root, but both crl_thigh__L and crl_thigh__R are relative to crl_hips__C), starting with the root point at 0.0, 0.0, 0.0. This is the ‘relative‘ pose.

The 3D pose can also be projected to 2D. The projection output is expressed in pixel values observed by the specific camera.