Content Classification for Live Streams
summary
A scene-classification filter runs on the GPU using ffmpeg TensorFlow backend and a MobileNetV2 model trained on ~80,000 frames, optimized to 224×224 input, and it detects categories like soccer and adult content. Benchmarks show an RTX 4000 handles 15 concurrent streams transcode+classify using ~40% GPU. The team ships training scripts, data and weights as open source model and training data, and the filter integrates into ffmpeg filter graphs and object replacement workflows.
AI-written summary · about
chapters (8)
- 0:00Intro
- 4:40Optimize Performance for GPUs
- 6:11Custom Al Filter - Training the Model
- 6:52Custom Al Filter - Pick a DNN backend
- 7:22Custom Al Filter Architecture
- 7:55Custom Al Filter - Benchmarking
- 9:09Next Steps - Chaining Multiple Filters
- 10:10Next Steps - Open Source Community
related talks
51mVideo Analytics for Football games
48mLeveraging the Power of C++ for Efficient Machine Learning on Embedded Devices
35mNNStreamer: Efficient and Agile On-Device AI Application Development with Neural-Net Pipelines
20mAI-Driven Real-Time Ball Tracking for Live Sports Streaming
53mVideo Object Detection at Scale
34mTesting Live Video Streams At Scale
30mTensorFlow for Mobile Heroes
27mOn challenges and considerations for real time AI processing