Enhanced human pose estimation via self-distilled and token-pruned transformer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Human Pose Estimation (HPE) is crucial for understanding human actions in images.
• Convolutional Neural Network (CNN) models have been traditionally used for HPE.
• Vision Transformer (ViT) models show improved performance but demand high computational resources.

Purpose of the Study:

• To develop an efficient human pose estimation method that reduces computational costs.
• To maintain high performance in HPE while minimizing model complexity.
• To introduce SPTPose, a novel approach leveraging self-distillation and token pruning.

Main Methods:

• Employed self-distillation techniques to transfer knowledge from a larger model to a smaller one.
• Utilized token pruning to remove redundant information within the Vision Transformer architecture.
• Developed the SPTPose model, incorporating these efficiency-enhancing strategies.

Main Results:

• SPTPose-B achieved a mean Average Precision (mAP) of 74.8% on the MSCOCO validation dataset.
• The model has only 13.2 million parameters, significantly reducing computational load.
• Achieved 4.7 GFLOPs, demonstrating high computational efficiency.

Conclusions:

• SPTPose offers a compelling solution for efficient and high-performance human pose estimation.
• The method successfully balances accuracy with reduced computational requirements.
• This work makes advanced HPE models more practical for resource-constrained environments.