MONOCULAR 3D BASED HUMAN POSE ESTIMATION WITH

REFINEMENT BLOCK AND SPECIAL LOSS FUNCTION

Abstract:

In this architecture, we present a 3D HPE by monocular. We use the multi-loss method that depends on 2D heatmaps and volumetric heatmaps and a refinement block to locate root-relative 3D human pose. Our approach takes 2D heatmaps and volumetric heatmaps as features to compute loss and combine the loss from relative 3D location to generate the total loss. The model can learn the 2D location and 3D location jointly and focus on the root-relative 3D position in the camera coordinate.

Network architecture:

We especially focus on the volumetric heatmaps, and adopt combinational loss to calculate the total loss. In summary, we present an end-to-end approach based on multi-loss and refinement block to estimate 3D human pose from a sing-view RGB image.

Experiment Results:

These are the visualization of our model. Top: RGB images of inputs. Middle: The 3D skeleton of ground truth. Bottom: The predicted 3D skeleton from out model.

Conclusion:

In this architecture, we propose a refinement block which is composed by the fully-connected layer and two residual blocks based on deep learning to estimate root-relative 3D human pose in camera coordinate. We utilize the loss from 2D heatmaps and volumetric heatmaps and 3D skeleton to calculate the total loss.

Made by ùÖöºÂ