Transformer-based 3D pose estimation pipeline with modular SmoothNet integration for animation generation.

Area of Science:

• Computer Vision
• 3D Human Pose Estimation
• Animation Generation

Background:

• Traditional 3D human pose estimation methods suffer from high computational costs and temporal instability.
• Existing techniques often struggle with accurate spatial prediction and smooth animation sequences.

Purpose of the Study:

• To develop an efficient and temporally stable 3D human pose estimation pipeline.
• To improve the quality of 3D human animation generation by reducing jitter.
• To evaluate the effectiveness of a Transformer-based backbone and SmoothNet integration.

Main Methods:

• A Transformer-based backbone utilizing ResNet for feature extraction and Transformer blocks for global dependency encoding.
• Integration of SmoothNet as a modular post-processing component for temporal smoothing of pose sequences.
• Evaluation using standard metrics (MPJPE, PA-MPJPE, MPVE, acceleration error) on H3WB, AGORA, and AthletePose3D datasets.

Main Results:

• The proposed pipeline demonstrates improved spatial prediction and significant reduction in temporal jitter.
• A fairness-controlled protocol confirmed the effectiveness of the backbone and SmoothNet independently.
• Successful animation quality validation in Blender and Unreal Engine (UE).

Conclusions:

• The Transformer-based pipeline with SmoothNet effectively addresses computational cost and temporal instability in 3D human pose estimation.
• The method offers a robust solution for generating smoother and more accurate 3D human animations.
• The approach is validated across multiple datasets and practical animation software.