Validation of deep learning-based markerless 3D pose estimation

Deep learning-based approaches to markerless 3D pose estimation are being adopted by researchers in psychology and neuroscience at an unprecedented rate. Yet many of these tools remain unvalidated. Here, we report on the validation of one increasingly popular tool (DeepLabCut) against simultaneous measurements obtained from a reference measurement system (Fastrak) with well-known performance characteristics. Our results confirm close (mm range) agreement between the two, indicating that under specific circumstances deep learning-based approaches can match more traditional motion tracking methods. Although more work needs to be done to determine their specific performance characteristics and limitations, this study should help build confidence within the research community using these new tools.

Automated summary

This study reports on the validation of a deep learning-based pose estimation tool (DeepLabCut) against a traditional measurement system (Fastrak). The results show that, under specific circumstances, deep learning-based approaches can achieve similar performance to traditional motion tracking methods.