HPnet: Hybrid Parallel Network for Human Pose Estimation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Hybrid models combining convolution and transformer networks show promise for human pose estimation.
• Existing sequential hybrid models suffer from mutual module conflict, hindering precise keypoint localization.
• This conflict leads to inconsistent performance between overall accuracy and high-precision localization.

Purpose of the Study:

• To develop a novel hybrid parallel network for human pose estimation.
• To alleviate the mutual conflict observed in sequential hybrid models.
• To improve both semantic representation and high-precision keypoint localization.

Main Methods:

• Designed a hybrid parallel network by processing convolution and self-attention modules concurrently.
• Utilized the self-attention branch for modeling long-range dependencies and semantic representation.
• Employed the convolution branch for local sensitivity and high-precision localization.
• Introduced a cross-branches attention module to gate features from both branches.

Main Results:

• The hybrid parallel network achieved 75.6% and 75.4% AP on the COCO validation and test-dev sets, respectively.
• Demonstrated consistent performance in both higher-precision keypoint localization and overall accuracy.
• The proposed model performs comparably to existing state-of-the-art human pose estimation methods.

Conclusions:

• The hybrid parallel network effectively leverages complementary capabilities of convolution and self-attention modules.
• Parallel processing mitigates module conflict, leading to improved and consistent human pose estimation performance.
• The proposed architecture offers a competitive alternative to current state-of-the-art approaches.