LCFFNet: A Lightweight Cross-scale Feature Fusion Network for human pose estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Human pose estimation is crucial in computer vision but faces challenges in balancing model efficiency and accuracy.
• Existing models often have high computational costs and large parameter volumes.

Purpose of the Study:

• To develop a lightweight yet accurate human pose estimation network.
• To address the trade-off between model complexity and performance in pose estimation.

Main Methods:

• Introduced the Lightweight Cross-scale Feature Fusion Network (LCFFNet).
• LCFFNet comprises a Lightweight HRNet-Like (LHRNet) network with Dynamic Multi-scale Convolution Basic (DMSC-Basic) blocks, a Cross-Resolution-Aware Semantics Module (CRASM), and an Adapt Feature Fusion Module (AFFM).
• Employed dynamic multi-scale convolution for reduced parameters and enhanced feature extraction, CRASM for fusing multi-scale features, and AFFM for adaptive feature fusion.

Main Results:

• Achieved 74.2% AP on MSCOCO 2017, 89.9% PCKh@0.5 on MPII, and 66.9% AP on Crowdpose datasets.
• Reduced model parameters by 89.0% and computational complexity by 87.5% compared to HRNet.
• Demonstrated performance comparable to large models and superior to state-of-the-art lightweight networks.

Conclusions:

• LCFFNet effectively balances accuracy and efficiency in human pose estimation.
• The proposed network offers a lightweight solution without compromising performance.
• LCFFNet represents a significant advancement in efficient and accurate human pose estimation.