Boggart: Towards General-Purpose Acceleration of Retrospective Video Analytics
summary
Boggart builds a preprocessor that uses classical CV—background subtraction, morphological ops, connected components, keypoints and descriptors—to produce motion trajectories and then profiles clusters at query time to decide which frames to run CNNs on. On 96 hours of public footage it runs the full query model on 5% of frames yet achieves 93% for a 90% target (counts and bounding boxes), outperforms model-specific Focus, and pre-processes on CPU without GPUs.
AI-written summary · about
chapters (4)
- 0:00Start
- 4:12Preprocessing Requirements
- 8:58Query Execution: New Techniques
- 12:27Comparison to Model-Specific Preprocessing
related talks
15mVideo Analytics with Zero-streaming Cameras
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
27mAnalysis of sports scenes using computer vision models
27mDevelop Computer Vision Apps With Go
40m"Introduction to Face Processing with Computer Vision|"
12mCascade models in computer vision to boost accuracy and performance
23mAVA: Towards Agentic Video Analytics with Vision Language Models