Popup: reconstructing 3D video using particle filtering to aggregate crowd responses
summary
A pipeline reconstructs 3D vehicle position and orientation from 2D video by collecting crowd-drawn dimension lines and aggregating them across frames with particle filtering, borrowing annotations from neighboring frames to exploit temporal content diversity. Workers can self-filter to skip hard annotations; this lowers annotation time by ~16% and, on KITTI (17 videos, 10 frames each, five workers/frame via AMT), yields a reported ~33% reduction in relative position error versus a strong baseline.
AI-written summary · about
chapters (8)
- 0:00Introduction to research
- 1:53Challenges in 3D reconstruction
- 4:43Accuracy issues in annotations
- 6:55Particle filtering method
- 9:26Experiment and filtering results
- 12:20Performance evaluation
- 14:29Summary and future work
- 15:10Q&A session
related talks
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
41mSocial Behavior Prediction from Video Observations
26mFrom Panoramic Vision to Lightweight 3D Worlds for Semantic Visual Tracking
41mComputer Vision for Remote AR
13mFlyMeThrough: Human-AI Collaborative 3D Indoor Mapping with Commodity Drones
17mData Challenges with 3D Computer Vision #bbuzz
35mDeep learning for 3d human motion capture
15mBoggart: Towards General-Purpose Acceleration of Retrospective Video Analytics