CGFusionFormer: Exploring Compact Spatial Representation for Robust 3D Human Pose Estimation with Low Computation...

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Transformer-based methods excel at 2D-to-3D human pose lifting.
• Challenges remain in handling low-quality 2D joint data and high computational demands.

Purpose of the Study:

• Introduce CGFusionFormer to address limitations in current 2D-to-3D human pose estimation.
• Improve robustness to poor 2D joint quality and reduce computational overhead.

Main Methods:

• Propose Compact Spatial Representation (CSR) for robust local feature generation.
• Utilize a Hybrid Adaptive Fusion module combining spatial and frequency domain features.
• Implement CGFusionFormer using a PoseFormer-like transformer backbone.

Main Results:

• Achieve superior accuracy-efficiency trade-off on Human3.6M and MPI-INF-3DHP benchmarks.
• Attain 47.6 mm MPJPE with 71.3 MFLOPs on Human3.6M (40% computation reduction vs. PoseFormerV2).
• Reach 97.9% PCK, 78.5% AUC, and 27.2 mm MPJPE on MPI-INF-3DHP, matching top performance.

Conclusions:

• CGFusionFormer offers a highly accurate and computationally efficient solution for 3D human pose estimation.
• The proposed CSR and fusion module effectively handle noisy 2D inputs and reduce computational load.
• Demonstrates significant improvements over existing transformer-based approaches, especially with short receptive fields.