LabelAR: A Spatial Guidance Interface for Fast Computer Vision Image Collection
summary
LabelAR uses an AR phone app to place persistent 3D bounding volumes, auto-project 2D boxes as images are captured, and guide users to 24 viewpoints so images are captured automatically. A user study with 12 participants produces 36 trained Faster R‑CNN models; LabelAR collects 24 images per object set, labels 118.5 objects on average, and achieves comparable detection to costly post‑hoc labeling in about 6.1 minutes versus 16 minutes, and ≈11× better accuracy than a single overlaid box.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:11Existing labeling methods
- 2:59Introducing LabelAR
- 4:42Design benefits and trade-offs
- 6:11User study setup
- 8:20Model training and evaluation
- 10:51Results and performance
- 12:37Performance factors analysis
- 14:23Next steps and future work
- 15:05Conclusion and Q&A
related talks
41mAI applications in autonomous driving
41mFusing AI with AR: Custom Markers
39mA Real use case of practical AR-AI couple: Size my luggage
37mA Real use case of practical AR-AI couple - Size my luggage
29mLabeling Tools are Great, but What About Quality Checks?
39mNow smile (also in 3D)! Exploring AR, ML and Camera related APIs on Android
30mStarting deep learning projects without enough labelled data
30mIn the service of the history. AI in archivistics.