JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

Lightweight cattle pose estimation with fusion of reparameterization and an attention mechanism

  • 0School of Engineering, Dali University, Dali, China.
Plos One +

|

August 7, 2024

|

August 7, 2024

View abstract on PubMed

Summary

This summary is machine-generated.

A new lightweight cattle pose estimation network reduces complexity and improves speed without heatmaps. This method enhances accuracy and efficiency for livestock applications.

Area Of Science

  • Computer Vision
  • Animal Science
  • Machine Learning

Background

  • Heatmap-based cattle pose estimation methods exhibit high network complexity and slow detection speeds.
  • Existing methods struggle with complex scenarios, necessitating more efficient approaches.

Purpose Of The Study

  • To propose an end-to-end, lightweight cattle pose estimation network that overcomes the limitations of heatmap-based methods.
  • To enhance network performance through a reparameterized network and attention mechanisms for complex scenarios.

Main Methods

  • Developed the EfficientRepBiPAN (Efficient Representation Bi-Directional Progressive Attention Network) module for the neck network to capture multi-scale features and reduce redundancy.
  • Incorporated a 3D parameterless SimAM (Similarity-based Attention Mechanism) into the backbone to capture directional and positional information.
  • Evaluated the model using a dataset of 6846 images.

Main Results

  • The proposed network achieved a 4.3% increase in average accuracy (at OKS = 0.5) compared to the baseline method.
  • Reduced floating-point computations by 1.0 G and parameters by 0.16 M.
  • Outperformed existing models like HRNet and YOLOv5-pose, improving AP0.5 by at least 0.4% while reducing parameters and computation.

Conclusions

  • The proposed lightweight network offers a harmonious balance between accuracy and efficiency in cattle pose estimation.
  • This method provides a valuable theoretical reference for pose estimation in various livestock industries.
  • The EfficientRepBiPAN and SimAM modules contribute significantly to improved performance and reduced computational load.