Large-scale Capture Of Hidden Fluorescent Labels For Training Generalizable Markerless Motion Capture Models

We present two pipelines for turning fluorescent labels into trained landmark trackers. The serial labeling pipeline enables reliable tracking of specific landmarks of interest, while the parallel labeling pipeline enables tracking of an arbitrary number of automatically selected landmarks. Our approach to generating labeled images is based on applying hidden fluorescent fiducials to body regions of interest, and we develop several innovations to optimize it for generating high-quality labels at large scale. We demonstrate the efficacy of our approach by using it to train a deep learning model for tracking the hand of the mouse and evaluate its accuracy on challenging image data.