Designing Videogames to Crowdsource Accelerometer Data Annotation for Activity Recognition Research
summary
Two web-based human computation games are built to let casual players label raw wrist accelerometer traces: a rhythm game Mobots and a puzzle aligner Signaligner. Pilots on Amazon Mechanical Turk run with 100 players per game using labeled non-dominant wrist data (walking, running, sedentary, sleep), splitting validation and unknown sets. Players reach 89% accuracy in Mobots and 99% accuracy in Signaligner; results show puzzles beat rhythm games and can correct algorithm errors.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:36Outline
- 1:05What is Activity Recognition
- 2:39Scenario
- 4:55Pilot Study
- 5:27Results
- 6:23Signal Laner
- 9:07Summary
related talks
08mSenseCollect: We Need Efficient Ways to Collect On-body Sensor-based Human Activity Data!
29mHow the NFL builds computer vision training datasets at scale
08mAgents and Robots: The Human Effort Putting AI Into Production
28mDiscussion: Humans As Sensors
23mCustom-made games with machine learning and big data
17mSession: Accelerating Decision Making From Ideation To Live Ops – A Data-Driven Approach With Game..
15mWhat Gamee learned about hyper-casual games from 3 billion gameplays across 70 games
20mAn Innovative Approach to Labeling Ground Truth in Speech