Using pose estimation to identify regions and points on natural history specimens

Area of Science:

• Biodiversity informatics
• Computational biology
• Digital specimen analysis

Background:

• Mobilizing digitized biological specimens for research requires efficient methods for extracting phenotypic data.
• Current methods for phenotypic measurement are often low-throughput, limiting the use of large digital collections.

Purpose of the Study:

• To evaluate a Deep Learning-based pose estimation approach for high-throughput phenotypic data extraction from digitized biological specimens.
• To demonstrate the utility of pose estimation for analyzing avian plumage coloration and snail shell morphometrics.

Main Methods:

• A pose estimation model was trained to place point labels on specimen images.
• The model was applied to avian specimens for plumage color analysis and to snail shells for morphometric shape analysis.
• Performance was validated by comparing model-derived measurements with human-expert measurements.

Main Results:

• The pose estimation approach achieved over 95% accuracy in labeling avian images and accurately measured plumage coloration.
• For snail shells, over 95% of landmarks were accurately placed, reliably capturing ecotype-specific shape variations.
• Measurements derived from pose estimation showed high correlation with human-based measurements.

Conclusions:

• Deep Learning-based pose estimation offers a high-throughput, accurate solution for extracting point-based phenotypic measurements from digitized biodiversity data.
• This method has the potential to significantly advance the mobilization and utilization of large-scale biological image datasets.
• Guidelines are provided for implementing pose estimation on similar biodiversity datasets.