jetson-inference

jetson-inference is an instructional project and library for deploying deep-learning vision models on NVIDIA Jetson devices. It uses TensorRT to run optimized networks from C++ or Python, while PyTorch is used for model training and transfer learning. The library provides interfaces for image classification, object detection, semantic segmentation, pose estimation, action recognition, background removal, and monocular depth estimation. The README includes C++ and Python examples for processing images and live camera feeds, as well as tutorials for camera streaming, WebRTC web applications, Flask REST services, and Plotly dashboards. ROS and ROS2 integration is provided through the related ros_deep_learning project. The Hello AI World material covers running inference entirely on Jetson, collecting datasets, retraining models, and deploying the resulting models with TensorRT. The same PyTorch training code can also be used on a PC, server, or cloud instance with an NVIDIA discrete GPU. The project includes pre-trained models for ImageNet-style recognition, COCO object detection, traffic and dash-cam scenarios, face detection, Cityscapes and Pascal VOC segmentation, body and hand pose estimation, and Kinetics action recognition. Source builds, Docker containers, and JetPack setup instructions are provided, with JetPack 6 support noted for Orin devices. The C++ and Python APIs can be linked into external projects through libjetson-inference and libjetson-utils.