Annotation of SMPLX models

Annotation pipeline for annotating pose and camera parameters.

The annotation pipeline contains mainly two modules.

  • Pose curation module.
  • Camera parameters curation module.

We use the off-the-shelf method like TCMR and ProHMR to get initial SMPLX pose. These methods take the image sequences as an input and predict the SMPLX pose parameters. These methods often fail on Indian classical dance scenarios either due to the lack of such training data or lack of proper approaches to handle such complex poses.

In pose curation module, we curate the wrong pose predictions using an orthographic projection of the mesh in the image. We use the open-source tool Blender for this task.

Pose annotation pipeline in Blender

Once poses are curated in step 1, we still can’t project these mesh to the 2D image because we don’t have the camera parameters. Hence we develop PyQT framework to curate the camera parameters. We use the same weak perspective camera model as used by TCMR.

Curation of camera parameters.