LiftPose3D

LiftPose3D reconstructs three-dimensional (3D) animal poses from single two-dimensional (2D) camera views to enable kinematic studies in laboratory animals.

Key Features:

• Single-View 3D Reconstruction: Produces accurate 3D pose estimates from a single 2D camera view without requiring multiple synchronized cameras.
• Deep Learning Lifting: Leverages advanced deep learning networks to transform 2D pose estimates into 3D models.
• Replacement of Multi-View Triangulation: Addresses limitations of traditional multi-view triangulation by eliminating the need for complex calibration protocols.
• Robustness to Occlusion and Camera Angles: Maintains accurate pose estimation under varying camera angles and partial occlusions typical of freely moving animals.
• Cross-Species Validation: Demonstrated on diverse laboratory animal models including flies, mice, rats, and macaques.

Scientific Applications:

• Kinematic Studies: Provides 3D pose data for quantitative analysis of movement and biomechanics in laboratory animals.
• Behavioral Quantification in Freely Moving Animals: Enables tracking and pose estimation when body parts are occluded or animals move freely.
• Cross-Species Comparative Analysis: Supports comparative studies across species such as flies, mice, rats, and macaques using a unified single-camera approach.

Methodology:

Transforms 2D pose estimates into 3D reconstructions using advanced deep learning networks, avoiding reliance on multi-view triangulation, synchronized camera arrays, or complex calibration.