Reproduce “ODIN: Automated Drift Detection and Recovery in Video Analytics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Abhijit Suprem, Joy Arulraj, Calton Pu, Joao Ferreira
ODIN is an automated framework that detects and recovers from concept drift in video analytics, adapting to changing visual data by deploying specialized models. This system, developed by researchers at Georgia Institute of Technology and the University of Sao Paulo, achieved 6x higher throughput, 2x higher object detection accuracy, and 6x smaller memory footprint compared to static systems.
Key insights
- Learning a low-dimensional, distance-preserving latent representation via Dual-Adversarial GAN (DA-GAN) effectively overcomes the "curse of dimensionality" for drift detection in high-dimensional visual data.
- Using "density bands" in the latent space allows for robust statistical detection of concept drift by identifying new data distributions that fall outside established high-density regions.
- Automatically training and deploying specialized, lightweight models (e.g., YOLO-LITE) via a student-teacher approach enables efficient recovery from drift while maintaining high accuracy for specific data distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.