DeepLabCut: markerless pose estimation of user-defined body parts with deep learning

DeepLabCut is a toolbox for markerless pose estimation aimed at quantifying behavior, a crucial but labor-intensive task in neuroscience. Traditional motor-control studies attach reflective markers to subjects, but markers are intrusive and require deciding their number and location in advance. Instead, DeepLabCut uses transfer learning with deep neural networks to track user-defined body parts directly from video, achieving strong results from only a small amount of labeled training data.

The authors demonstrate the framework's versatility by tracking a variety of body parts across multiple species and a broad range of behaviors. Remarkably, even when only about 200 frames are labeled, the algorithm achieves tracking performance on held-out test frames comparable to human accuracy. This combination of minimal labeling effort and human-level precision made DeepLabCut a widely useful tool for behavioral quantification.