Palleon: A Runtime System for Efficient Video Processing toward Dynamic Class Skew
summary
Palleon is a runtime system that detects dynamic class skew from prediction streams, uses early sampling and windowed switch detection, applies a Bayesian filter (with posterior rescoring and a fast-threshold pass) and performs online model selection from a cached edge-cloud duplication model bank. Evaluated on real videos (6–24 minutes, up to 45 class-skew switches, 2–3.5 classes on average), Palleon yields up to 6.7x energy savings and 7.9x latency reduction while preserving target accuracy.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:21Motivation
- 1:09Temporal Locality: Class Skew
- 2:08Class Skew: Class Cardinality
- 2:38Class Skew: Visual Separability
- 3:18Challenges
- 4:39System Overview
- 5:15ABLE
- 7:29Bayesian Filter
- 9:00Separability-Aware Model Selection
- 11:36Evaluation
related talks
33mArchitecting Scalable Multi-Modal Video Search
15mBoggart: Towards General-Purpose Acceleration of Retrospective Video Analytics
25mOptimal Video Compression Using Pixel Shift Tracking
16mSwift: Adaptive Video Streaming with Layered Neural Codecs
17mYour Player is Smarter Than You Think: Self-Learning in Adaptive Streaming
14mRegion-based Content Enhancement for Efficient Video Analytics at the Edge
20mBuilding a Professional Video Editor on the Cloud, Powered by Machine Learning
35mLarge-Scale Video Intelligence