posenet idea

Sports App based on posenet / openPose

<iframe width="560" height="315" src="https://www.youtube.com/embed/pff0i-vICoI" frameborder="0" allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>

Summary

for most sports use cases, a high level of accuracy and frame-by-frame stability is needed. Carnegie Mellon University's (CMU) <span class="text-highlight">openPose</span> seems to provide satisfactory results, but their commercial license doesn not allow use in anything sports-related.

We would like to discuss the feasibility of tuning google's open source model, <span class="text-highlight">posenet</span> to match that of openPose's. Realtime predictions are considered secondary - accuracy of pose matching is the KPI.

Our current prototype is focused on Golf, but we are currently also discussing non sports-related possibilities.

Golf Study

key poses

• 9 key poses identified from two main viewpoints

accuracy with posenet

• insufficient accuracy (critical poses not captured accurately)
• unstable predictions (prediction keypoints jumps around)

App Study

MVP UI (subject to change)

Project Status

MVP