From Forest to Zoo: Great Ape Behavior Recognition with ChimpBehave

Area of Science:

• Primate ethology
• Computer vision
• Machine learning

Background:

• Automated behavior recognition is vital for primate research and conservation.
• Manual video annotation is time-consuming, hindering large-scale studies.
• Machine learning models struggle with varied data collection contexts and annotations.

Purpose of the Study:

• Introduce ChimpBehave, a novel dataset for chimpanzee behavior recognition.
• Enable research on cross-dataset generalization for machine learning models.
• Establish performance baselines for automated behavior recognition models.

Main Methods:

• Developed ChimpBehave dataset: ~2h 20min video of zoo-housed chimpanzees.
• Annotated with bounding boxes and fine-grained locomotive behaviors.
• Aligned behavior classes with the PanAf dataset for cross-dataset generalization studies.

Main Results:

• Benchmarked state-of-the-art video-based and skeleton-based action recognition models.
• Established performance baselines for within-dataset and cross-dataset evaluations.
• Highlighted strengths and limitations of different model architectures.

Conclusions:

• ChimpBehave dataset facilitates research on automated behavior recognition across diverse settings.
• Findings provide insights into applying machine learning for primate behavior analysis.
• Dataset, models, and code are publicly available for further research.