OpenLabCluster: active learning based clustering and classification of animal behaviors based on kinematic body...

Area of Science:

• Ethology and Behavioral Biology
• Computational Biology and Bioinformatics
• Machine Learning in Biological Sciences

Background:

• Quantifying natural animal behavior from video is crucial for ethological studies.
• Markerless pose estimation infers body keypoints, but organizing these sequences into behaviors is challenging.
• Existing methods often require extensive human annotations, leading to suboptimal performance due to limited representative samples.

Purpose of the Study:

• To introduce a general toolset, OpenLabCluster, for efficient organization and interpretation of keypoint sequences into behavioral categories.
• To reduce the need for human annotations in behavioral classification.
• To improve the accuracy and speed of behavioral analysis across various animal species.

Main Methods:

• OpenLabCluster clusters temporal keypoint segments in a latent space.
• An Active Learning (AL) approach refines clusters and classifies them into behavioral states.
• The AL approach strategically selects representative segments for annotation, informing the clustering and classification of all segments.

Main Results:

• Demonstrated OpenLabCluster's effectiveness on four diverse datasets across different animal species.
• Showcased boosted clustering and classification performance compared to existing methods.
• Achieved superior results even when compared to methods using full annotation.

Conclusions:

• OpenLabCluster significantly reduces human annotation requirements for behavioral analysis.
• The toolset enables faster and more accurate organization and classification of behavioral segments.
• OpenLabCluster is an open-source, interactive graphic interface designed for scientific use, incorporating user feedback.