PigeonPose: An integrated pipeline for markerless 3D tracking of pigeon behavior

Behavioral science, a multidisciplinary field spanning neuroscience, psychology, and ecology, aims to understand individual, collective, and social behaviors along with their neural and genetic correlates. Nevertheless, quantifying naturalistic behaviors remains challenging due to time-consuming manual methods that are prone to errors and inconsistencies. Integrating techniques from physics, computer science, and mathematics has improved this process, particularly in pose estimation. Several tools for pose estimation exist, but they have been mainly designed for standard animal models such as rodents and their adoption can be technically complex. However, given that birds are increasingly popular in neuroscientific research there is an obvious lack of analysis tools to detect unique avian behaviors such as head bobbing or flying. Here we address the above-mentioned challenges and provide a scalable solution for markerless 3D tracking during long recording sessions of up to 2 hours. Our pipeline leverages DeepLabCut and Anipose for behavioral tracking and includes a software toolkit for synchronizing multi-camera video capture with external hardware, a Python module for data compression, analysis, triangulation, and behavior analysis. We also provide a large labeled dataset for pigeon tracking and pre-trained models with robust out-of-the-box performance. In documenting our experimental setup and analysis pipeline, we offer guidance on video recording, compression, and data management for future users. Finally, we demonstrate the tracking results of single and multiple freely moving pigeons in an open field setup and in a restricted operant chamber, highlighting the potential for new data modalities and analysis opportunities from 3D pose tracking.