// the find
sherlockchou86/VideoPipe
A cross-platform video structuring (video analysis) framework. If you find it helpful, please give it a star: ) 跨平台的视频结构化（视频分析）框架，觉得有帮助的请给个星星 : )
VideoPipe is a C++ pipeline framework for video analysis tasks — object detection, tracking, behavior analysis, face recognition, license plate reading. It sits in the same space as NVIDIA DeepStream but without the vendor lock-in, targeting surveillance and traffic monitoring scenarios where you bring your own models.
The node-based pipeline API is genuinely clean — attaching nodes with `attach_to({previous_node})` and getting fan-out for free is the right abstraction for this domain. Backend agnosticism is real: you can run the same pipeline with OpenCV::DNN on CPU or swap in TensorRT/ONNX Runtime/PaddleInference without touching the pipeline structure. The multimodal LLM integration (Ollama/vLLM/OpenAI-compatible) added in 2025 is interesting — lets you drop a scene-understanding node anywhere in the pipeline without special wiring. The 40+ working sample programs lower the 'does this actually work?' question significantly.
Ubuntu-only with a hard dependency on GStreamer 1.14.5 — 'cross-platform' in the name oversells it; Windows and macOS are not supported, and the tested ARM targets are Jetson/RK3588 specifically. Performance is explicitly rated 'medium' compared to DeepStream, which matters a lot in production surveillance deployments where you might be running 16+ streams per box. No Python bindings means the data science crowd who'd actually integrate the LLM nodes can't use it without writing C++. The test data and models live on Google Drive and Baidu Drive with no checksums — that's a brittle setup for reproducible builds.