AnimalRTPose: Faster cross-species real-time animal pose estimation

Xin Wu ¹, Lianming Wang ², Jipeng Huang ³

⁰School of Physics, Northeast Normal University, Renmin Street 5268, Changchun, Jilin, 130024, China. Electronic address: wux024@nenu.edu.cn.

Neural Networks : the Official Journal of the International Neural Network Society +

|

June 14, 2025

View abstract on PubMed

Summary

AnimalRTPose is a novel one-stage model for real-time animal pose estimation across species. It achieves high accuracy and speed, outperforming existing models for behavior analysis.

Area Of Science

Computer Vision
Animal Behavior Analysis
Machine Learning

Background

Analyzing complex animal behaviors is crucial but challenging due to diverse morphologies and environments.
Existing computer vision tools struggle with real-time pose estimation across different species.

Purpose Of The Study

Introduce AnimalRTPose, a one-stage model for cross-species, real-time animal pose estimation.
Address limitations in current pose estimation methods for diverse animal subjects and conditions.

Main Methods

Developed AnimalRTPose, a one-stage model utilizing CSPNeXt backbone for feature extraction.
Integrated channel attention mechanism (CAM) and spatial pyramid pooling (SPP) for robust feature representation.
Employed an efficient multi-scale feature fusion module for dynamic balancing of local and global information.

Main Results

AnimalRTPose demonstrates superior performance in real-time animal pose estimation, surpassing existing one-stage models.
Achieved high frame rates (e.g., 1111 FPS on NVIDIA A800) and efficient throughput on edge devices (e.g., 196 FPS on NVIDIA Jetson AGX Orin).
The model supports single-animal, multi-animal, cross-species, and few-shot scenarios with high accuracy and robustness.

Conclusions

AnimalRTPose offers a scalable and versatile solution for real-time animal pose estimation.
Its high performance makes it applicable for various scenarios, including real-time animal behavior monitoring.
The model's architecture effectively handles diverse animal morphologies and environmental complexities.

Keywords:

Animal pose estimation Cross-species Lightweight network Real-time