CoVA: Exploiting Compressed-Domain Analysis to Accelerate Video Analytics
summary
CoVA builds a three-stage pipeline that finds moving-object tracks from motion vectors using a learned model called Plumnet, picks a minimal set of frames via dependency-aware frame selection to avoid expensive GOP decoding, decodes only those frames for full DNN labels, and propagates labels along tracks. Implementation uses block motion vectors and an encoder–decoder temporal net, cuts decoding by ~80% and DNN inference by ~99%, and yields 4.8× end-to-end throughput with ~12% geometric-mean drop in binary-query accuracy.
AI-written summary · about
chapters (8)
- 0:00Introduction to challenges
- 2:28Prior works overview
- 5:32Limitations and Kova concept
- 7:12Track detection stage
- 10:43Frame selection
- 13:33Label propagation
- 14:21Evaluation and results
- 17:23Conclusion
related talks
33mArchitecting Scalable Multi-Modal Video Search
35mLarge-Scale Video Intelligence
08mReconstructing 3D from Compressed Video: An AV1-Based SfM Pipeline
35mFootball video analysis using Deep Learning
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
25mOptimal Video Compression Using Pixel Shift Tracking
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
22mAI Enhanced GPU Video Coding: Achieving Joint High Compression Efficiency and Throughput