jetson-inference is a C++ and Python library for deploying deep learning inference on NVIDIA Jetson embedded devices using TensorRT optimization.
The library addresses the challenge of running neural networks efficiently on resource-constrained edge hardware. It wraps TensorRT to execute optimized inference directly from C++ or Python, supporting multiple vision tasks including image classification, object detection, semantic segmentation, pose estimation, and action recognition. The approach combines pre-trained models with the ability to perform transfer learning and deploy custom models trained in PyTorch, all executable onboard the Jetson device itself.
Developers working on embedded computer vision projects targeting Jetson hardware should consider this tool, particularly those building robotics applications, video analytics systems, or real-time vision pipelines. The project provides examples for streaming from live camera feeds, creating WebRTC-based webapps, and integrating with ROS and ROS2. It supports multiple Jetson platforms and includes a structured tutorial called Hello AI World that guides users from running inference through collecting datasets and training custom models. The library includes pre-trained models and handles the full pipeline from model training to optimized deployment.
The project maintains active development with ongoing support for new Jetson hardware platforms and JetPack versions. Documentation is comprehensive, covering system setup, building from source, and Docker container deployment. The codebase includes extensive code examples demonstrating inference workflows and integration patterns. The project regularly incorporates new capabilities, as evidenced by expanding support for generative AI and large language model tutorials alongside the core vision primitives.
