Paper

HandAugment: A Simple Data Augmentation Method for Depth-Based 3D Hand Pose Estimation

Problem

Depth-based 3D hand pose estimation remains difficult because of viewpoint variation, self-occlusion, self-similarity between fingers, and the limited availability of accurately labeled training data.

Method

This work combines two ideas: a two-stage neural network pipeline for improving hand region extraction from depth images, and a data augmentation strategy based on MANO-driven synthetic depth generation and real-synthetic blending.

Results

The method achieved first place in the depth-based 3D hand pose estimation task of the HANDS 2019 challenge and showed strong improvements over baseline hand pose estimators.

Focus

  • Depth-based 3D hand pose estimation
  • Better hand region extraction from noisy input patches
  • Synthetic data augmentation for improved generalization

Updated: