Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Current human pose estimation datasets lack sufficient scale and diversity.
• Realistic human sensing systems require large, accurately annotated 3D pose data.

Purpose of the Study:

• Introduce Human3.6M, a large-scale dataset for 3D human pose estimation.
• Provide diverse human activities, synchronized multi-modal data, and 3D body scans.
• Establish benchmarks and evaluation scenarios for advanced human sensing.

Main Methods:

• Acquired 3.6 million accurate 3D human poses from 11 subjects across 4 viewpoints.
• Collected synchronized image, motion capture, and time-of-flight (depth) data.
• Developed large-scale statistical models and controlled mixed reality evaluation scenarios.

Main Results:

• Achieved a 20% performance improvement using the full Human3.6M training set compared to existing datasets.
• Demonstrated the dataset's diversity and potential for future research.
• Established evaluation baselines for human pose estimation.

Conclusions:

• Human3.6M significantly advances the state-of-the-art in 3D human pose estimation.
• The dataset's scale and diversity enable training of more robust and accurate human sensing systems.
• Further research leveraging complex models with this dataset promises substantial improvements.