MadEye: Boosting Live Video Analytics Accuracy with Adaptive Camera Configurations
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
MadEye is a camera–server system that actively reorients commodity PTZ cameras, runs a fast search over rotations/zooms, and uses on‑camera lightweight detectors with a knowledge‑distillation selection step to pick the frame sent for backend inference. It reports 2.9–25.7% accuracy gains over best fixed orientations or the same accuracy with 2–3.7× lower resource costs. The prototype sweeps many orientations per timestep, trains on-server and pushes distilled weights to cameras, and yields consistent improvements across workloads.
AI-written summary · about
chapters (8)
- 0:00MadEye introduction and goals
- 2:23Camera orientation importance
- 5:16PTZ optimization challenges
- 8:07MadEye’s adaptive exploration
- 9:33Orientation selection for accuracy
- 12:02On-camera approximation models
- 14:30MadEye workflow and evaluation
- 16:15Conclusion and key takeaways
related talks
20mBuilding a Personalized, Auto-Calibrating Eye Tracker from User Interactions
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
20mEye-tracking and AI: Unveiling the Limitations and Synergies
28mVision, Big Data and AI
19mEyeScout: Active Eye Tracking for Position and Movement Independent Gaze Interaction with Large ...
29mApplied Custom Vision
42mDeep Learning for Automatic Basketball Video Production
06mMemX: An Attention-Aware Smart Eyewear System for Personalized Moment Auto-capture