A model that works in a notebook is not yet a vision system. The last step is to measure how good it really is, make it fast enough to be useful, and put it where people can use it, whether that is a server, a phone or camera, or a web page.
Evaluation starts with the right metrics. For detection and segmentation that means intersection over union (IoU) and mean average precision (mAP), which score both whether you found an object and how accurately you outlined it. Optimization then shrinks and speeds up the model: ONNX is a portable format for moving a trained model between frameworks and runtimes, and TensorRT compiles it for fast inference on NVIDIA GPUs.
Finally, deployment is about where the model runs: behind an API on a server, on an edge device close to the camera, or directly in the browser. Watch the lessons in order, then put everything together in the Projects.
What mean average precision measures, how it is calculated, and how it is used when evaluating a YOLOv3 detector.
Converting a PyTorch model to the ONNX format and running it in a web page, with a comparison of browser-based inference and a backend server.
A short explainer on TensorRT, the NVIDIA library built to run machine learning models quickly.
An introduction to running computer vision models on edge devices, close to where the images are captured.