Augmented reality
Crowdsourced data for gesture recognition improved model accuracy by 14%.

Training data for AI and machine learning in augmented reality

About case:
About case:
Industry and use case
AR, gesture recognition
DATA
15,000 videos
Project duration
21 days
Industry and use case
AR, gesture recognition
DATA
15,000 videos
Project duration
21 days
Challenge:
Challenge:
The customer is developing a smart TV box with an RGB camera that recognizes users' gestures. To improve the quality of gesture recognition, the company asked the Training Data team to collect and annotate 15,000 videos with 14 various gestures. The main challenge was to mine videos that mirror actual users' gestures in real life when a user set up the TV box above the TV.
Solution:
Solution:
We crowdsourced video collection, while our team created detailed requirements for the video and organized strict quality control.
Outcomes:
Outcomes:
We collected and annotated videos filmed by 7,000 different people
Gesture recognition accuracy was improved by 14%
Feedback
We came up with an idea, but we didn't have precise requirements. The CEO and CTO advised us and built the entire data collection process. The team independently developed the video collection requirements and crowdsourced the video collection. Very flexible team that allowed us to experiment and refine our needs to improve the final product!
Alex P. Product Owner