HDPose: Post-Hierarchical Diffusion with Conditioning for 3D Human Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Monocular 3D human pose estimation (HPE) addresses the 3D-2D projection ill-posed problem but faces challenges with depth ambiguity and occlusions.
• Diffusion model-based approaches (DDPM) reconstruct 3D poses from noisy inputs, often using 2D keypoints or context encoders for spatial-temporal information.

Purpose of the Study:

• To develop a novel monocular 3D HPE method that achieves rapid convergence and high accuracy.
• To improve upon existing diffusion model-based approaches that may suffer from slow convergence or suboptimal performance.

Main Methods:

• Proposed HDPose, a diffusion model-based approach for monocular 3D HPE.
• Aggregated spatial and temporal information within a hierarchical denoising model structure.
• Investigated various condition structures, identifying the post-hierarchical structure as optimal.

Main Results:

• HDPose demonstrated rapid convergence and accurate 3D pose prediction.
• Achieved competitive performance against state-of-the-art models on Human3.6M and MPI-INF-3DHP datasets.
• The model is considerably more lightweight compared to existing methods.

Conclusions:

• The proposed hierarchical aggregation of spatial-temporal information enhances diffusion model performance for monocular 3D HPE.
• HDPose offers a promising solution for accurate, efficient, and lightweight 3D human pose estimation.
• The findings suggest potential for further advancements in real-time and resource-constrained HPE applications.