Learning-Deep-Learning

DensePose: Dense Human Pose Estimation In The Wild

September 2020

tl;dr: Annotating human surface correspondence on RGB images and adapt Mask RCNN for prediction.

Overall impression

The paper proposed DensePose COCO dataset, and establishes dense correspondences between an RGB image and a surface-based representation of the human body.

Key ideas

• Partitioning and UV parameterization of the human body.
  • Since the human body has a complicated structure, we break it into multiple independent pieces and parameterize each piece using a local two-dimensional coordinate system, that identifies the position of any node on this surface part, based on unwrapping of the SMPL model.
• Sample points for annotation: roughly equidistant points obtained via k-means and request annotators to bring these points in correspondence with the surface. Note that the sample points are selected in the image domain, not surface domain.
• Knowledge distillation: Only a randomly chosen subset of image pixels per training sample is annotated. These sparse correspondence is to train a teacher model and then in-paint the supervision signal on the rest of the image mask.
  • Much better results can be obtained by in-painting the values of supervision signals that were not originally annotated. –> See Data distillation.
• Evaluation: via geodesic distance (actually euclidean distance in UV space).
• Architecture:
  • Modified Mask RCNN for this purpose
  • Regress part class (24 + 1), and then 24 regressor heads to predict (u, v).
  • Cross-cascading multitask learning: The output of one head feeds into a different task. The additional guidance from keypoint prediction branch boosts the dense pose prediction a lot.

Technical details

• The user interface for 3D annotation is well designed. Once the target point is annotated, the point is displayed on all rendered images simultaneously. This avoid manually rotating the surface.
• Accuracy of human annotators: generate images with GT by rendering human CAD models. This avoids different annotators to annotate the same object to get “consensus” GT.
  • Errors are small on small surface parts with distinctive features such as face, hands and feet. For large uniform areas the annotator errors can get larger.

Notes

• Domain adaptation can be achieved by Gradient Reversal.