DeepPoseKit

DeepPoseKit performs fast, robust 2D pose estimation of user-defined animal keypoints using deep learning for quantitative animal behavior analysis.

Key Features:

• 2D pose estimation: Performs 2D pose estimation of user-defined keypoints in animals from images and videos.
• Stacked DenseNet: Implements a multi-scale deep-learning architecture (Stacked DenseNet) optimized for animal pose estimation.
• Implementation: Implemented in Python and built on the TensorFlow and Keras frameworks.
• GPU-based peak detection: Uses a fast GPU-based peak-detection algorithm to localize keypoints with subpixel precision.
• High throughput: Processes images and videos at speeds reported to exceed twice the throughput of prior methods while maintaining accuracy.
• Robustness: Validated in challenging laboratory and field scenarios and capable of handling groups of interacting individuals.

Scientific Applications:

• Quantitative behavioral analysis: Extracts precise keypoint data for quantitative measurement of animal behaviors.
• Neuroscience experiments: Provides behavioral metrics for studies linking neural activity to movement and posture.
• Ecology and field studies: Enables pose-based analysis of animals in field settings and ecological research.
• Social and interaction studies: Supports analysis of groups of interacting individuals and social behaviors.

Methodology:

Implemented in Python using TensorFlow and Keras, employing a multi-scale Stacked DenseNet architecture and a GPU-based peak-detection algorithm for subpixel keypoint localization from images and videos.