OpenApePose, a database of annotated ape photographs for pose estimation

Area of Science:

• Primatology and Computer Vision
• Behavioral Ecology
• Machine Learning for Animal Tracking

Background:

• Non-human apes are of significant scientific interest due to their close relationship with humans.
• Understanding complex ape behavior is crucial and can be advanced by video-based pose tracking.
• High-quality annotated datasets are essential for effective pose tracking in animals.

Purpose of the Study:

• To introduce OpenApePose, a novel public dataset for ape pose tracking.
• To evaluate the effectiveness of specialized ape pose tracking models.
• To demonstrate the importance of tailored datasets for animal behavior analysis.

Main Methods:

• Creation of OpenApePose: a dataset of 71,868 annotated photographs of six ape species.
• Annotation includes 16 body landmarks for each ape in naturalistic settings.
• Training and evaluation of a deep neural network (HRNet-W48) on the OpenApePose dataset.

Main Results:

• A deep network trained on OpenApePose significantly outperforms models trained on monkey (OpenMonkeyPose) and human (COCO) datasets for ape pose tracking.
• The specialized ape model achieves tracking accuracy comparable to models trained for their respective taxa.
• Excluding one species during training improved tracking performance for that specific held-out ape species.

Conclusions:

• Large, specialized datasets are critical for developing robust animal pose tracking systems.
• OpenApePose is a valuable resource that significantly enhances the ability to track ape behavior.
• The findings underscore the utility of domain-specific data in advancing computer vision applications for wildlife research.