ESTABLISHING TRAINING PARAMETERS FOR A DEEP NEURAL NETWORK TO ASSESS 2D, FRONTAL PLANE KINEMATICS

Authors

Caleb D. Johnson, Harvard Medical School, Department of Physical Medicine and Rehabilitation, Cambridge, MA, USAFollow
Lauren K. Sara, Harvard Medical School, Department of Physical Medicine and Rehabilitation, Cambridge, MA, USA
Taylor Kofton, Harvard Medical School, Department of Physical Medicine and Rehabilitation, Cambridge, MA, USA
William Marshall, Harvard Medical School, Department of Physical Medicine and Rehabilitation, Cambridge, MA, USA
Julie M. Hughes, Military Performance Division, U.S. Army Research Institute of Environmental Medicine, Natick MA, USA
Stephen A. Foulis, Military Performance Division, U.S. Army Research Institute of Environmental Medicine, Natick MA, USA
Irene S. Davis, Harvard Medical School, Department of Physical Medicine and Rehabilitation, Cambridge, MA, USA

Category

Technology/equipment

Document Type

Paper

Abstract

The purpose of this study was to establish the optimal training parameters to assess frontal plane, 2D kinematics using DeepLabCut. DeepLabCut is an open-source platform that allows the user to train neural networks for customized feature detection in 2D videos. Deep neural networks were trained using frontal plane videos from 41 participants who completed single- and double-leg drop landings. Networks were trained with an increasing number of training iterations (25-250k) and training frames (200-800). Our results indicate that a minimum of 175k training iterations and 400 training frames were adequate for stable network performance (training/test errors= 2.8/3.7 pixels).

Recommended Citation

Johnson, Caleb D.; Sara, Lauren K.; Kofton, Taylor; Marshall, William; Hughes, Julie M.; Foulis, Stephen A.; and Davis, Irene S. (2022) "ESTABLISHING TRAINING PARAMETERS FOR A DEEP NEURAL NETWORK TO ASSESS 2D, FRONTAL PLANE KINEMATICS," ISBS Proceedings Archive: Vol. 40: Iss. 1, Article 75.
Available at: https://commons.nmu.edu/isbs/vol40/iss1/75