PolarBearVidID: A Video-Based Re-Identification Benchmark Dataset for Polar Bears

Area of Science:

• Zoological research
• Computer vision
• Machine learning

Background:

• Automated monitoring systems are vital for zoological institutions studying animal behavior.
• Re-identification of individual animals across multiple cameras is a critical processing step.
• Deep learning, particularly video-based methods, shows promise for re-identification tasks, leveraging animal movement as a feature.

Purpose of the Study:

• To introduce PolarBearVidID, the first video-based re-identification dataset for a non-human species.
• To address challenges in zoo environments like variable lighting, occlusions, and low resolutions.
• To train and test a video-based re-identification approach on this novel dataset.

Main Methods:

• Creation of an extensively annotated dataset (PolarBearVidID) with 13 polar bears across 1431 sequences (138,363 images).
• Filming polar bears in unconstrained poses and diverse lighting conditions, unlike typical human re-identification datasets.
• Development and evaluation of a video-based re-identification algorithm.

Main Results:

• The trained re-identification approach achieved a rank-1 accuracy of 96.6% on the PolarBearVidID dataset.
• Demonstrated the effectiveness of using animal movement as a distinguishing feature for re-identification.
• Successfully addressed challenges posed by unconstrained environments in zoological settings.

Conclusions:

• PolarBearVidID is a valuable resource for advancing video-based re-identification in zoological studies.
• Animal movement is a robust characteristic feature for accurate re-identification in challenging conditions.
• Deep learning approaches utilizing video data can significantly enhance automated monitoring systems in zoos.