My slides on this paper for the CS 637 Computer Vision II class.

Problem Statement

  • Aims at mapping human pixels from an image to 3D surface of a template human body
  • Task involves object detection, pose estimation, part and instance segmentation

3D Human Template

  • SMPL – A Skinned Multi-Person Linear Model
  • Per part UV Unwrapping to regress UV coordinates

Overall Pipeline

Good Techniques used in the paper:

  • Cascading
  • Averaging background
  • Perform evaluation at multiple scales
  • Using a Teacher Network (FCNN) to boost annotations
Information sharing between branches, additional information from keypoint prediction task helps improve results.
Using COCO person segmentation masks to averageout background
To boost annotations; Train a FCNN on single person crops, calculate loss on labelled points and ignore unlabelled.

The Dataset

  • Image to surface correspondences on 50k persons from COCO dataset
  • Total 5 million correspondences => ~100 per person
  • Max 14 correspondences per body part
The annotation interface to human labellers

Labelling Errors from annotators

  • Large errors ~8cm in labelling on torso and thighs due to lack of distintictive points.

Results

Application: Texture Transfer

Links: