jetson-inference
Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.
A guided tutorial library for running computer vision AI on NVIDIA Jetson edge devices, covering object detection, segmentation, pose estimation, and training custom models.
NVIDIA Jetson is a family of small, energy-efficient computing boards designed to run AI tasks at the edge, meaning on a physical device rather than in a remote data center. jetson-inference is a guided library and tutorial series, called Hello AI World, that shows you how to run computer vision AI models on Jetson hardware. It is aimed at developers who want to add visual AI to robots, cameras, or embedded systems without sending data to the cloud.
The library is built around TensorRT, NVIDIA's software for running AI models efficiently on their GPUs. It provides ready-to-use modules for several common vision tasks: identifying what is in an image (classification), drawing boxes around specific objects (detection), coloring each pixel by what it belongs to (segmentation), estimating where a person's joints are positioned (pose estimation), identifying what action a person is performing (action recognition), removing background from images, and estimating depth from a single camera. These modules can be called from either C++ or Python code.
The Hello AI World tutorial walks through the full workflow: setting up a Jetson device, running pre-trained AI models on live camera input, collecting your own image datasets, training your own classification or detection models using PyTorch, and then deploying those custom models back to the device with TensorRT. Video walkthroughs from an NVIDIA AI certification course are linked in the README for people who prefer watching over reading.
The library also includes tools for building web-based demos using WebRTC, which allows streaming video from the Jetson to a browser in real time. It works with ROS and ROS2, the standard software frameworks used in robotics.
A companion resource called Jetson AI Lab covers newer topics such as large language models, vision language models, and vision transformers running on the more powerful Jetson Orin hardware. The two resources share the same community but Jetson AI Lab is a separate site focused on generative AI workloads rather than traditional computer vision.
Where it fits
- Run real-time object detection on a Jetson camera feed to identify and track people or items on a robot.
- Train a custom image classifier with your own photos using PyTorch and deploy it to a Jetson board with TensorRT.
- Build a live semantic segmentation pipeline that labels each camera pixel by category on an embedded device.
- Stream a computer vision demo from a Jetson to a browser in real time using the built-in WebRTC tools.