FlyMeThrough: Human-AI Collaborative 3D Indoor Mapping with Commodity Drones
summary
A drone-to-map pipeline converts consumer drone RGB videos into 3D reconstructions using Metashape photogrammetry, lets users click objects on frames, propagates masks with video segmentation, and localizes POIs via depth-guided ray casting into the mesh to produce 3D bounding boxes. Evaluated across 12 indoor sites (reconstruction succeeds in 11), flights finish within 10 minutes, annotations take ~15 minutes, and 91% of 158 annotations become 3D boxes with 71 judged high quality.
AI-written summary · about
chapters (8)
- 0:00Project motivation
- 1:18System overview
- 2:063D reconstruction
- 3:55POI annotation
- 6:08System evaluation
- 8:06Insights and applications
- 9:31Conclusion and Q&A
- 10:58Extended Q&A
related talks
13mRobust and Accurate Augmentation of Municipal Environments
20mInteractive Room Capture on 3D-Aware Mobile Devices
36mDeep Learning - Segmentation, Identification and Classification
45m(Deep) Learning to Fly
41mComputer Vision for Remote AR
57mDrones & AI - What's all the buzz about ?
35mDeep learning for 3d human motion capture
12mReality Capture & Visualization for Treehouse Design