// the find
pipeless-ai/pipeless
An open-source computer vision framework to build and deploy apps in minutes
Pipeless is a video stream processing framework that wraps GStreamer and ONNX/OpenVINO/TensorRT runtimes so you can write pre/process/post Python hooks without wiring up the pipeline plumbing yourself. The core is Rust, but user code is Python. It targets developers who want to ship a YOLO-based detection app without becoming a GStreamer expert.
The stage abstraction is genuinely clean — each stage is a directory with hook files, which keeps processing logic self-contained and composable across streams. Supporting ONNX, OpenVINO, TensorRT, and CoreML through a single URL-based model loading interface saves real integration work. Dynamic per-stream pipeline configuration via REST API without restarting the process is useful for multi-camera deployments. Docker images for CUDA and TensorRT variants are pre-built and tested, which is usually the painful part of getting GPU inference running.
Last commit was May 2024 and the hosted agents product linked in the README appears to be the commercial pivot — this repo looks abandoned in favor of that. GStreamer 1.20.3 is a hard pinned requirement that will fight you on any modern Linux distro shipping 1.22 or 1.24. Python hooks running inside a Rust process via FFI means debugging is painful — stack traces from hook failures are opaque. The 'multi-language support' claim is aspirational; looking at the source, Rust hooks exist but are not documented with examples, so you're effectively Python-only.