3D Bird Reconstruction: a Dataset, Model, and Shape Recovery from a Single View.

Area of Science:

• Ethology
• Computer Vision
• Computational Neuroscience

Background:

• Automated animal pose estimation is crucial for studying animal behavior and neuroscience.
• Current methods struggle with occluded social animals like birds, limiting research scope.

Purpose of the Study:

• To develop a robust method for capturing avian pose and shape, addressing occlusion challenges.
• To enable accurate recovery of bird postures from single-view images.

Main Methods:

• Introduced a novel model and multi-view optimization approach for avian pose and shape.
• Developed a pipeline for keypoint, mask, pose, and shape regression from single views.
• Collected multi-view keypoint and mask annotations from 15 social birds in an outdoor aviary.

Main Results:

• Successfully captured the unique shape and pose space of live birds.
• Recovered accurate avian postures from single-view estimations.
• Created the extensive Penn Aviary Dataset for future research.

Conclusions:

• The new method significantly improves automated pose estimation for social birds.
• This work provides valuable tools and datasets for advancing avian behavior and neuroscience research.