PoseCoach

Cover
Great pose
Bad pose
Great pose

PoseCoach

Many people are struggling to give good presentations. A large part of it is not even related to the content, but rather _ how _ it is being presented. Studies have shown high correlations between (A) rather subconscious factors such as posture, facial expression, or tone of voice, and (B) perceived quality of the presentation.

These "A"-factors can be controlled by the presenter, which is why professional speech coaches pay special attention on them during rehearsal/training sessions. We therefore wanted to build software that helps people become better presenters through a real-time evaluation of these factors.

During the last few years, machine learning has become incredibly powerful and our goal was to build a system that automatically detects bad posture and informs the user about it in real-time.

We used open-source software to pre-process and display information through the browser (HTML/CSS/JS, ml5js, Python/Django, bulma) and a pre-trained network for pose detection (Tensorflow PoseNet). As always, more (variable) training data is the ultimate factor for precision (we used about 10 minutes of selfmade videos). However, even with little input, our software is able to produce reasonably accurate results for a MVP. Similarly, further features can be added (speech analysis, facial expressions, etc.).