Data Challenges with 3D Computer Vision #bbuzz
summary
Automated 3D capture from Kinect/Project Tango, stereo rigs and LIDAR produces huge raw volumes: a 9 MP 8‑bit color + 16‑bit depth frame is ~43 MB, 10 fps → ~36 TB/day. Practical pipeline uses implicit voxel/TSDF stored in octrees backed by hash maps, enabling ~0.4 µs voxel access and real‑time integration at ~2 fps on current rigs. Drones and cars demonstrate city‑scale meshes; a 100×100 m area at 3 cm resolution compresses to ~70 MB, but search, standardization and production‑grade visualization remain unsolved.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:003D Computer Vision Challenges
- 5:44Digital Perception: From 2D to a 3D Modell
- 7:473D Perception From single-shot to full 3D model
- 8:20Digital Perception: Data Challenges
- 9:07How can we...
- 10:23Storage for 3D Modelling Representing the 3D world as voxels
- 15:31Conclusion
related talks
41mComputer Vision for Remote AR
26mFrom Panoramic Vision to Lightweight 3D Worlds for Semantic Visual Tracking
28mVision, Big Data and AI
20mInteractive Room Capture on 3D-Aware Mobile Devices
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
17mBuilding Training Data for Autonomous Driving
28mData Curation: Transforming Bytes into AI Gold
35mDeep learning for 3d human motion capture