FiGO

Code for paper - "FiGO: Fine-Grained Query Optimization for Video Analytics" (SIGMOD'2022).

Installation

conda env create -f environment.yml

Hardware

• TITAN Xp + CUDA 11.6 Ideas can be tested as well by using cached inference results.

Example Run

The easiest way to evaluate the idea of FiGO with other systems is to use cached results (avoid extra inference time).

# run q1 end-to-end execution with all system implementations
python experiment/end_to_end.py --cache-path $PWD/demo-res --query q1

Accuracy (F-1 score) and query processing time should be avaiable at res/q1/res.txt.

Currently Supported Systems

• FiGO
• MS-Filter: Filter-based approach
• ME-Coarse: Coarse-grained optimization that entire video uses a single model
• MC: Model cascade approach
• Naive: Run the most accuracy model on all frames. Used for accuracy evaluation.

Documentation

• Every system implements a scheduler, which does both query optimization and execution. The scheduler will take a loader to load decoded frames of a video.
• Query directory contains queries implemented by using provided scheduler and predicates.
• Experiment folder contains an end-to-end script to run all systems and collect results.