Deep Imputation for Skeleton data (DISK) for behavioral science.

Area of Science:

• Animal behavior research
• Biomechanical analysis
• Machine learning applications

Background:

• Quantitative animal kinematics measurements are crucial for behavior studies.
• Tracking errors in pose estimation and motion capture systems often render experimental data unusable.
• Manual annotation of missing data is labor-intensive and impractical for large datasets.

Purpose of the Study:

• To develop a deep learning method for imputing missing animal pose data without manual annotations.
• To improve the analysis of animal behavior experiments by recovering unusable tracking data.
• To create a versatile tool applicable to various tracking methods and downstream analyses.

Main Methods:

• Developed Deep Imputation for Skeleton data (DISK), a deep learning model.
• Leveraged keypoint dependencies and dynamics to impute missing skeleton data.
• Validated DISK on seven diverse animal skeletons, including multi-animal scenarios.

Main Results:

• DISK successfully imputes missing tracking data across various animal models and multi-animal setups.
• Imputed data enabled detection of more motion episodes (e.g., steps) and increased statistical robustness.
• The method learned meaningful data representations, capturing underlying animal actions.

Conclusions:

• DISK offers a powerful, automated solution for handling missing data in animal pose estimation.
• The imputation package enhances the reliability and scope of animal behavior research.
• DISK is a valuable, open-source tool for researchers utilizing motion capture and pose estimation techniques.