// the find
dusty-nv/jetson-inference
Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.
NVIDIA's official tutorial repo for running deep learning inference on Jetson devices using TensorRT, with Python and C++ bindings for classification, detection, segmentation, pose estimation, and more. It's aimed at embedded ML practitioners who own Jetson hardware and want to get vision models running with hardware acceleration without writing TensorRT boilerplate from scratch. The documentation is genuinely good — step-by-step, with video walkthroughs and pre-trained models that download automatically.
The abstraction layer over TensorRT is well-designed — imageNet/detectNet/segNet classes hide engine building, FP16/INT8 calibration, and memory management behind a clean API that works identically from C++ and Python. Pre-trained model zoo covers the common use cases (SSD-MobileNet, FCN-ResNet18 segmentation at multiple resolutions, pose estimation) so you can validate the pipeline before touching your own models. The WebRTC server integration is a genuine differentiator — streaming inference results to a browser over WebRTC without third-party dependencies is non-trivial and it's already wired up. JetPack 6 / Orin support was added, so the repo isn't stale on the latest hardware generation.
The model selection is frozen in 2020-era architectures — ResNet-18, SSD-MobileNet-v2, AlexNet-based segmentation. There's no YOLO v8/v9, no ViT-based detectors, nothing from the last four years of the COCO leaderboard. You can bring your own ONNX model but the docs on that path are thin compared to the pre-trained model path. The INT8 calibration support uses a random data calibrator by default, which produces measurably worse accuracy than a real calibration dataset — this is mentioned nowhere in the beginner docs and will surprise anyone who deploys INT8 and wonders why it performs worse than expected. The C++ API is well-documented but the Python bindings documentation is auto-generated from C++ and the type annotations are missing, making IDE integration painful.