[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-dusty-nv--jetson-inference":3,"tool-dusty-nv--jetson-inference":61},[4,18,26,36,44,53],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":17},4358,"openclaw","openclaw\u002Fopenclaw","OpenClaw 是一款专为个人打造的本地化 AI 助手，旨在让你在自己的设备上拥有完全可控的智能伙伴。它打破了传统 AI 助手局限于特定网页或应用的束缚，能够直接接入你日常使用的各类通讯渠道，包括微信、WhatsApp、Telegram、Discord、iMessage 等数十种平台。无论你在哪个聊天软件中发送消息，OpenClaw 都能即时响应，甚至支持在 macOS、iOS 和 Android 设备上进行语音交互，并提供实时的画布渲染功能供你操控。\n\n这款工具主要解决了用户对数据隐私、响应速度以及“始终在线”体验的需求。通过将 AI 部署在本地，用户无需依赖云端服务即可享受快速、私密的智能辅助，真正实现了“你的数据，你做主”。其独特的技术亮点在于强大的网关架构，将控制平面与核心助手分离，确保跨平台通信的流畅性与扩展性。\n\nOpenClaw 非常适合希望构建个性化工作流的技术爱好者、开发者，以及注重隐私保护且不愿被单一生态绑定的普通用户。只要具备基础的终端操作能力（支持 macOS、Linux 及 Windows WSL2），即可通过简单的命令行引导完成部署。如果你渴望拥有一个懂你",349277,3,"2026-04-06T06:32:30",[13,14,15,16],"Agent","开发框架","图像","数据工具","ready",{"id":19,"name":20,"github_repo":21,"description_zh":22,"stars":23,"difficulty_score":10,"last_commit_at":24,"category_tags":25,"status":17},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,"2026-04-05T11:01:52",[14,15,13],{"id":27,"name":28,"github_repo":29,"description_zh":30,"stars":31,"difficulty_score":32,"last_commit_at":33,"category_tags":34,"status":17},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",154349,2,"2026-04-13T23:32:16",[14,13,35],"语言模型",{"id":37,"name":38,"github_repo":39,"description_zh":40,"stars":41,"difficulty_score":32,"last_commit_at":42,"category_tags":43,"status":17},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",108322,"2026-04-10T11:39:34",[14,15,13],{"id":45,"name":46,"github_repo":47,"description_zh":48,"stars":49,"difficulty_score":32,"last_commit_at":50,"category_tags":51,"status":17},6121,"gemini-cli","google-gemini\u002Fgemini-cli","gemini-cli 是一款由谷歌推出的开源 AI 命令行工具，它将强大的 Gemini 大模型能力直接集成到用户的终端环境中。对于习惯在命令行工作的开发者而言，它提供了一条从输入提示词到获取模型响应的最短路径，无需切换窗口即可享受智能辅助。\n\n这款工具主要解决了开发过程中频繁上下文切换的痛点，让用户能在熟悉的终端界面内直接完成代码理解、生成、调试以及自动化运维任务。无论是查询大型代码库、根据草图生成应用，还是执行复杂的 Git 操作，gemini-cli 都能通过自然语言指令高效处理。\n\n它特别适合广大软件工程师、DevOps 人员及技术研究人员使用。其核心亮点包括支持高达 100 万 token 的超长上下文窗口，具备出色的逻辑推理能力；内置 Google 搜索、文件操作及 Shell 命令执行等实用工具；更独特的是，它支持 MCP（模型上下文协议），允许用户灵活扩展自定义集成，连接如图像生成等外部能力。此外，个人谷歌账号即可享受免费的额度支持，且项目基于 Apache 2.0 协议完全开源，是提升终端工作效率的理想助手。",100752,"2026-04-10T01:20:03",[52,13,15,14],"插件",{"id":54,"name":55,"github_repo":56,"description_zh":57,"stars":58,"difficulty_score":32,"last_commit_at":59,"category_tags":60,"status":17},4721,"markitdown","microsoft\u002Fmarkitdown","MarkItDown 是一款由微软 AutoGen 团队打造的轻量级 Python 工具，专为将各类文件高效转换为 Markdown 格式而设计。它支持 PDF、Word、Excel、PPT、图片（含 OCR）、音频（含语音转录）、HTML 乃至 YouTube 链接等多种格式的解析，能够精准提取文档中的标题、列表、表格和链接等关键结构信息。\n\n在人工智能应用日益普及的今天，大语言模型（LLM）虽擅长处理文本，却难以直接读取复杂的二进制办公文档。MarkItDown 恰好解决了这一痛点，它将非结构化或半结构化的文件转化为模型“原生理解”且 Token 效率极高的 Markdown 格式，成为连接本地文件与 AI 分析 pipeline 的理想桥梁。此外，它还提供了 MCP（模型上下文协议）服务器，可无缝集成到 Claude Desktop 等 LLM 应用中。\n\n这款工具特别适合开发者、数据科学家及 AI 研究人员使用，尤其是那些需要构建文档检索增强生成（RAG）系统、进行批量文本分析或希望让 AI 助手直接“阅读”本地文件的用户。虽然生成的内容也具备一定可读性，但其核心优势在于为机器",93400,"2026-04-06T19:52:38",[52,14],{"id":62,"github_repo":63,"name":64,"description_en":65,"description_zh":66,"ai_summary_zh":67,"readme_en":68,"readme_zh":69,"quickstart_zh":70,"use_case_zh":71,"hero_image_url":72,"owner_login":73,"owner_name":74,"owner_avatar_url":75,"owner_bio":76,"owner_company":77,"owner_location":78,"owner_email":79,"owner_twitter":78,"owner_website":80,"owner_url":81,"languages":82,"stars":122,"forks":123,"last_commit_at":124,"license":125,"difficulty_score":126,"env_os":127,"env_gpu":128,"env_ram":129,"env_deps":130,"category_tags":138,"github_topics":141,"view_count":32,"oss_zip_url":78,"oss_zip_packed_at":78,"status":17,"created_at":162,"updated_at":163,"faqs":164,"releases":194},7314,"dusty-nv\u002Fjetson-inference","jetson-inference","Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.","jetson-inference 是专为 NVIDIA Jetson 嵌入式设备打造的深度学习部署指南与工具库，旨在帮助用户轻松将复杂的视觉算法落地到边缘端。它解决了在资源受限的硬件上高效运行实时 AI 视觉任务的难题，让开发者无需从零构建底层优化代码，即可快速实现图像分类、物体检测、语义分割、姿态估计及动作识别等功能。\n\n这套工具非常适合嵌入式 AI 开发者、机器人工程师以及希望探索边缘计算的研究人员使用。无论是想通过命令行快速体验预训练模型，还是希望通过 Python 或 C++ 编写自定义程序，甚至是利用 PyTorch 进行迁移学习来训练专属模型，jetson-inference 都提供了详尽的教程和示例。\n\n其核心亮点在于深度集成了 NVIDIA TensorRT 推理引擎，能够自动对神经网络进行优化，从而在 Jetson 的 GPU 上发挥极致的性能与能效比。此外，项目不仅支持从本地摄像头实时推流分析，还涵盖了 WebRTC 网页应用开发及 ROS\u002FROS2 机器人系统对接，甚至已扩展至生成式 AI 与大语言模型的应用场景。对于想要动手实践\"Hello AI World\"、","jetson-inference 是专为 NVIDIA Jetson 嵌入式设备打造的深度学习部署指南与工具库，旨在帮助用户轻松将复杂的视觉算法落地到边缘端。它解决了在资源受限的硬件上高效运行实时 AI 视觉任务的难题，让开发者无需从零构建底层优化代码，即可快速实现图像分类、物体检测、语义分割、姿态估计及动作识别等功能。\n\n这套工具非常适合嵌入式 AI 开发者、机器人工程师以及希望探索边缘计算的研究人员使用。无论是想通过命令行快速体验预训练模型，还是希望通过 Python 或 C++ 编写自定义程序，甚至是利用 PyTorch 进行迁移学习来训练专属模型，jetson-inference 都提供了详尽的教程和示例。\n\n其核心亮点在于深度集成了 NVIDIA TensorRT 推理引擎，能够自动对神经网络进行优化，从而在 Jetson 的 GPU 上发挥极致的性能与能效比。此外，项目不仅支持从本地摄像头实时推流分析，还涵盖了 WebRTC 网页应用开发及 ROS\u002FROS2 机器人系统对接，甚至已扩展至生成式 AI 与大语言模型的应用场景。对于想要动手实践\"Hello AI World\"、将理论模型转化为实际智能应用的创作者而言，这是一个不可多得的入门与实践平台。","\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_a68fe911e22b.jpg\" width=\"100%\">\n\n# Deploying Deep Learning\nWelcome to our instructional guide for inference and realtime vision [DNN library](#api-reference) for **[NVIDIA Jetson](https:\u002F\u002Fdeveloper.nvidia.com\u002Fembedded-computing)** devices.  This project uses **[TensorRT](https:\u002F\u002Fdeveloper.nvidia.com\u002Ftensorrt)** to run optimized networks on GPUs from C++ or Python, and PyTorch for training models.\n\nSupported DNN vision primitives include [`imageNet`](docs\u002Fimagenet-console-2.md) for image classification, [`detectNet`](docs\u002Fdetectnet-console-2.md) for object detection, [`segNet`](docs\u002Fsegnet-console-2.md) for semantic segmentation, [`poseNet`](docs\u002Fposenet.md) for pose estimation, and [`actionNet`](docs\u002Factionnet.md) for action recognition.  Examples are provided for streaming from live camera feeds, making webapps with WebRTC, and support for ROS\u002FROS2.\n\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_8b9f587bb1c5.jpg\">\n\nFollow the **[Hello AI World](#hello-ai-world)** tutorial for running inference and transfer learning onboard your Jetson, including collecting your own datasets, training your own models with PyTorch, and deploying them with TensorRT.\n\n### Table of Contents\n\n* [Hello AI World](#hello-ai-world)\n* [Jetson AI Lab](#jetson-ai-lab)\n* [Video Walkthroughs](#video-walkthroughs)\n* [API Reference](#api-reference)\n* [Code Examples](#code-examples)\n* [Pre-Trained Models](#pre-trained-models)\n* [System Requirements](#recommended-system-requirements)\n* [Change Log](CHANGELOG.md)\n\n> >   JetPack 6 is now supported on Orin devices ([developer.nvidia.com\u002Fjetpack](https:\u002F\u002Fdeveloper.nvidia.com\u002Fembedded\u002Fjetpack)) \u003Cbr\u002F>\n> >   Check out the Generative AI and LLM tutorials on [Jetson AI Lab](https:\u002F\u002Fwww.jetson-ai-lab.com\u002F)! \u003Cbr\u002F>\n> >   See the [Change Log](CHANGELOG.md) for the latest updates and new features. \u003Cbr\u002F>\n\n## Hello AI World\n\nHello AI World can be run completely onboard your Jetson, including live inferencing with TensorRT and transfer learning with PyTorch.  For installation instructions, see [System Setup](#system-setup).  It's then recommended to start with the [Inference](#inference) section to familiarize yourself with the concepts, before diving into [Training](#training) your own models.\n\n#### System Setup\n\n* [Setting up Jetson with JetPack](docs\u002Fjetpack-setup-2.md)\n* [Running the Docker Container](docs\u002Faux-docker.md)\n* [Building the Project from Source](docs\u002Fbuilding-repo-2.md)\n\n#### Inference\n\n* [Image Classification](docs\u002Fimagenet-console-2.md)\n\t* [Using the ImageNet Program on Jetson](docs\u002Fimagenet-console-2.md)\n\t* [Coding Your Own Image Recognition Program (Python)](docs\u002Fimagenet-example-python-2.md)\n\t* [Coding Your Own Image Recognition Program (C++)](docs\u002Fimagenet-example-2.md)\n\t* [Running the Live Camera Recognition Demo](docs\u002Fimagenet-camera-2.md)\n\t* [Multi-Label Classification for Image Tagging](docs\u002Fimagenet-tagging.md)\n* [Object Detection](docs\u002Fdetectnet-console-2.md)\n\t* [Detecting Objects from Images](docs\u002Fdetectnet-console-2.md#detecting-objects-from-the-command-line)\n\t* [Running the Live Camera Detection Demo](docs\u002Fdetectnet-camera-2.md)\n\t* [Coding Your Own Object Detection Program](docs\u002Fdetectnet-example-2.md)\n\t* [Using TAO Detection Models](docs\u002Fdetectnet-tao.md)\n\t* [Object Tracking on Video](docs\u002Fdetectnet-tracking.md)\n* [Semantic Segmentation](docs\u002Fsegnet-console-2.md)\n\t* [Segmenting Images from the Command Line](docs\u002Fsegnet-console-2.md#segmenting-images-from-the-command-line)\n\t* [Running the Live Camera Segmentation Demo](docs\u002Fsegnet-camera-2.md)\n* [Pose Estimation](docs\u002Fposenet.md)\n* [Action Recognition](docs\u002Factionnet.md)\n* [Background Removal](docs\u002Fbackgroundnet.md)\n* [Monocular Depth](docs\u002Fdepthnet.md)\n\n#### Training\n\n* [Transfer Learning with PyTorch](docs\u002Fpytorch-transfer-learning.md)\n* Classification\u002FRecognition (ResNet-18)\n\t* [Re-training on the Cat\u002FDog Dataset](docs\u002Fpytorch-cat-dog.md)\n\t* [Re-training on the PlantCLEF Dataset](docs\u002Fpytorch-plants.md)\n\t* [Collecting your own Classification Datasets](docs\u002Fpytorch-collect.md)\n* Object Detection (SSD-Mobilenet)\n\t* [Re-training SSD-Mobilenet](docs\u002Fpytorch-ssd.md)\n\t* [Collecting your own Detection Datasets](docs\u002Fpytorch-collect-detection.md)\n\n#### WebApp Frameworks\n\n* [WebRTC Server](docs\u002Fwebrtc-server.md)\n* [HTML \u002F JavaScript](docs\u002Fwebrtc-html.md)\n* [Flask + REST](docs\u002Fwebrtc-flask.md)\n* [Plotly Dashboard](docs\u002Fwebrtc-dash.md)\n* [Recognizer (Interactive Training)](docs\u002Fwebrtc-recognizer.md)\n\n#### Appendix\n\n* [Camera Streaming and Multimedia](docs\u002Faux-streaming.md)\n* [Image Manipulation with CUDA](docs\u002Faux-image.md)\n* [DNN Inference Nodes for ROS\u002FROS2](https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fros_deep_learning)\n\n## Jetson AI Lab\n\n\u003Ca href=\"https:\u002F\u002Fwww.jetson-ai-lab.com\">\u003Cimg align=\"right\" width=\"200\" height=\"200\" src=\"https:\u002F\u002Fnvidia-ai-iot.github.io\u002Fjetson-generative-ai-playground\u002Fimages\u002FJON_Gen-AI-panels.png\">\u003C\u002Fa>\n\nThe [**Jetson AI Lab**](https:\u002F\u002Fwww.jetson-ai-lab.com) has additional tutorials on LLMs, Vision Transformers (ViT), and Vision Language Models (VLM) that run on Orin (and in some cases Xavier).  Check out some of these:\n\n\u003Ca href=\"https:\u002F\u002Fwww.jetson-ai-lab.com\u002Ftutorial_nanoowl.html\">\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_5dcd1c4bb5c7.gif\">\u003C\u002Fa>\n> [NanoOWL - Open Vocabulary Object Detection ViT](https:\u002F\u002Fwww.jetson-ai-lab.com\u002Ftutorial_nanoowl.html) (container: [`nanoowl`](\u002Fpackages\u002Fvit\u002Fnanoowl)) \n\n\u003Ca href=\"https:\u002F\u002Fyoutu.be\u002FX-OXxPiUTuU\">\u003Cimg width=\"600px\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_63f2240c52ae.gif\">\u003C\u002Fa>\n> [Live Llava on Jetson AGX Orin](https:\u002F\u002Fyoutu.be\u002FX-OXxPiUTuU) (container: [`local_llm`](\u002Fpackages\u002Fllm\u002Flocal_llm#live-llava)) \n\n\u003Ca href=\"https:\u002F\u002Fyoutu.be\u002FdRmAGGuupuE\">\u003Cimg width=\"600px\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_04332a9ffc81.jpg\">\u003C\u002Fa>\n> [Live Llava 2.0 - VILA + Multimodal NanoDB on Jetson Orin](https:\u002F\u002Fyoutu.be\u002FX-OXxPiUTuU) (container: [`local_llm`](\u002Fpackages\u002Fllm\u002Flocal_llm#live-llava)) \n\n\u003Ca href=\"https:\u002F\u002Fyoutu.be\u002FayqKpQNd1Jw\">\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_caf4cb22628b.gif\">\u003C\u002Fa>\n> [Realtime Multimodal VectorDB on NVIDIA Jetson](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=wzLHAgDxMjQ) (container: [`nanodb`](\u002Fpackages\u002Fvectordb\u002Fnanodb))  \n\n## Video Walkthroughs\n\nBelow are screencasts of Hello AI World that were recorded for the [Jetson AI Certification](https:\u002F\u002Fdeveloper.nvidia.com\u002Fembedded\u002Flearn\u002Fjetson-ai-certification-programs) course:\n\n| Description                                                                                                                                                                                                                                                                                                        | Video                                                                                                                                                                                                                                                 |\n|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QXIwdsyK7Rw&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=9\" target=\"_blank\">**Hello AI World Setup**\u003C\u002Fa>\u003Cbr\u002F>Download and run the Hello AI World container on Jetson Nano, test your camera feed, and see how to stream it over the network via RTP.                                     | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QXIwdsyK7Rw&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=9\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_setup.jpg width=\"750\">\u003C\u002Fa>               |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QatH8iF0Efk&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=10\" target=\"_blank\">**Image Classification Inference**\u003C\u002Fa>\u003Cbr\u002F>Code your own Python program for image classification using Jetson Nano and deep learning, then experiment with realtime classification on a live camera stream. | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QatH8iF0Efk&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=10\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_imagenet.jpg width=\"750\">\u003C\u002Fa>           |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=sN6aT9TpltU&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=11\" target=\"_blank\">**Training Image Classification Models**\u003C\u002Fa>\u003Cbr\u002F>Learn how to train image classification models with PyTorch onboard Jetson Nano, and collect your own classification datasets to create custom models.     | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=sN6aT9TpltU&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=11\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_imagenet_training.jpg width=\"750\">\u003C\u002Fa>  |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=obt60r8ZeB0&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=12\" target=\"_blank\">**Object Detection Inference**\u003C\u002Fa>\u003Cbr\u002F>Code your own Python program for object detection using Jetson Nano and deep learning, then experiment with realtime detection on a live camera stream.              | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=obt60r8ZeB0&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=12\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_detectnet.jpg width=\"750\">\u003C\u002Fa>          |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=2XMkPW_sIGg&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=13\" target=\"_blank\">**Training Object Detection Models**\u003C\u002Fa>\u003Cbr\u002F>Learn how to train object detection models with PyTorch onboard Jetson Nano, and collect your own detection datasets to create custom models.                  | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=2XMkPW_sIGg&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=13\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_detectnet_training.jpg width=\"750\">\u003C\u002Fa> |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=AQhkMLaB_fY&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=14\" target=\"_blank\">**Semantic Segmentation**\u003C\u002Fa>\u003Cbr\u002F>Experiment with fully-convolutional semantic segmentation networks on Jetson Nano, and run realtime segmentation on a live camera stream.                                 | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=AQhkMLaB_fY&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=14\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_segnet.jpg width=\"750\">\u003C\u002Fa>             |\n\n## API Reference\n\nBelow are links to reference documentation for the [C++](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Findex.html) and [Python](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.html) libraries from the repo:\n\n#### jetson-inference\n\n|                    | [C++](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__deepVision.html) | [Python](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html) |\n|--------------------|--------------|--------------|\n| Image Recognition  | [`imageNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__imageNet.html#classimageNet) | [`imageNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#imageNet) |\n| Object Detection   | [`detectNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__detectNet.html#classdetectNet) | [`detectNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#detectNet)\n| Segmentation       | [`segNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__segNet.html#classsegNet) | [`segNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#segNet) |\n| Pose Estimation    | [`poseNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__poseNet.html#classposeNet) | [`poseNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#poseNet) |\n| Action Recognition | [`actionNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__actionNet.html#classactionNet) | [`actionNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#actionNet) |\n| Background Removal | [`backgroundNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__backgroundNet.html#classbackgroundNet) | [`actionNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#backgroundNet) |\n| Monocular Depth    | [`depthNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__depthNet.html#classdepthNet) | [`depthNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#depthNet) |\n\n#### jetson-utils\n\n* [C++](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__util.html)\n* [Python](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.utils.html)\n\nThese libraries are able to be used in external projects by linking to `libjetson-inference` and `libjetson-utils`.\n\n## Code Examples\n\nIntroductory code walkthroughs of using the library are covered during these steps of the Hello AI World tutorial:\n\n* [Coding Your Own Image Recognition Program (Python)](docs\u002Fimagenet-example-python-2.md)\n* [Coding Your Own Image Recognition Program (C++)](docs\u002Fimagenet-example-2.md)\n\nAdditional C++ and Python samples for running the networks on images and live camera streams can be found here:\n\n|                   | C++              | Python             |\n|-------------------|---------------------|---------------------|\n|    Image Recognition  | [`imagenet.cpp`](examples\u002Fimagenet\u002Fimagenet.cpp) | [`imagenet.py`](python\u002Fexamples\u002Fimagenet.py) |\n|    Object Detection   | [`detectnet.cpp`](examples\u002Fdetectnet\u002Fdetectnet.cpp) | [`detectnet.py`](python\u002Fexamples\u002Fdetectnet.py) |\n|    Segmentation       | [`segnet.cpp`](examples\u002Fsegnet\u002Fsegnet.cpp) | [`segnet.py`](python\u002Fexamples\u002Fsegnet.py) |\n|    Pose Estimation    | [`posenet.cpp`](examples\u002Fposenet\u002Fposenet.cpp) | [`posenet.py`](python\u002Fexamples\u002Fposenet.py) |\n|    Action Recognition | [`actionnet.cpp`](examples\u002Factionnet\u002Factionnet.cpp) | [`actionnet.py`](python\u002Fexamples\u002Factionnet.py) |\n|    Background Removal | [`backgroundnet.cpp`](examples\u002Fbackgroundnet\u002Fbackgroundnet.cpp) | [`backgroundnet.py`](python\u002Fexamples\u002Fbackgroundnet.py) |\n|    Monocular Depth    | [`depthnet.cpp`](examples\u002Fdepthnet\u002Fsegnet.cpp) | [`depthnet.py`](python\u002Fexamples\u002Fdepthnet.py) |\n\n> **note**:  see the [Array Interfaces](docs\u002Faux-image.md#array-interfaces) section for using memory with other Python libraries (like Numpy, PyTorch, ect)\n\nThese examples will automatically be compiled while [Building the Project from Source](docs\u002Fbuilding-repo-2.md), and are able to run the pre-trained models listed below in addition to custom models provided by the user.  Launch each example with `--help` for usage info.\n\n## Pre-Trained Models\n\nThe project comes with a number of pre-trained models that are available to use and will be automatically downloaded:\n\n#### Image Recognition\n\n| Network       | CLI argument   | NetworkType enum |\n| --------------|----------------|------------------|\n| AlexNet       | `alexnet`      | `ALEXNET`        |\n| GoogleNet     | `googlenet`    | `GOOGLENET`      |\n| GoogleNet-12  | `googlenet-12` | `GOOGLENET_12`   |\n| ResNet-18     | `resnet-18`    | `RESNET_18`      |\n| ResNet-50     | `resnet-50`    | `RESNET_50`      |\n| ResNet-101    | `resnet-101`   | `RESNET_101`     |\n| ResNet-152    | `resnet-152`   | `RESNET_152`     |\n| VGG-16        | `vgg-16`       | `VGG-16`         |\n| VGG-19        | `vgg-19`       | `VGG-19`         |\n| Inception-v4  | `inception-v4` | `INCEPTION_V4`   |\n\n#### Object Detection\n\n| Model                   | CLI argument       | NetworkType enum   | Object classes       |\n| ------------------------|--------------------|--------------------|----------------------|\n| SSD-Mobilenet-v1        | `ssd-mobilenet-v1` | `SSD_MOBILENET_V1` | 91 ([COCO classes](..\u002Fdata\u002Fnetworks\u002Fssd_coco_labels.txt))     |\n| SSD-Mobilenet-v2        | `ssd-mobilenet-v2` | `SSD_MOBILENET_V2` | 91 ([COCO classes](..\u002Fdata\u002Fnetworks\u002Fssd_coco_labels.txt))     |\n| SSD-Inception-v2        | `ssd-inception-v2` | `SSD_INCEPTION_V2` | 91 ([COCO classes](..\u002Fdata\u002Fnetworks\u002Fssd_coco_labels.txt))     |\n| TAO PeopleNet           | `peoplenet`        | `PEOPLENET`        | person, bag, face    |\n| TAO PeopleNet (pruned)  | `peoplenet-pruned` | `PEOPLENET_PRUNED` | person, bag, face    |\n| TAO DashCamNet          | `dashcamnet`       | `DASHCAMNET`       | person, car, bike, sign |\n| TAO TrafficCamNet       | `trafficcamnet`    | `TRAFFICCAMNET`    | person, car, bike, sign | \n| TAO FaceDetect          | `facedetect`       | `FACEDETECT`       | face                 |\n\n\u003Cdetails>\n\u003Csummary>Legacy Detection Models\u003C\u002Fsummary>\n\n| Model                   | CLI argument       | NetworkType enum   | Object classes       |\n| ------------------------|--------------------|--------------------|----------------------|\n| DetectNet-COCO-Dog      | `coco-dog`         | `COCO_DOG`         | dogs                 |\n| DetectNet-COCO-Bottle   | `coco-bottle`      | `COCO_BOTTLE`      | bottles              |\n| DetectNet-COCO-Chair    | `coco-chair`       | `COCO_CHAIR`       | chairs               |\n| DetectNet-COCO-Airplane | `coco-airplane`    | `COCO_AIRPLANE`    | airplanes            |\n| ped-100                 | `pednet`           | `PEDNET`           | pedestrians          |\n| multiped-500            | `multiped`         | `PEDNET_MULTI`     | pedestrians, luggage |\n| facenet-120             | `facenet`          | `FACENET`          | faces                |\n\n\u003C\u002Fdetails>\n\n#### Semantic Segmentation\n\n| Dataset      | Resolution | CLI Argument | Accuracy | Jetson Nano | Jetson Xavier |\n|:------------:|:----------:|--------------|:--------:|:-----------:|:-------------:|\n| [Cityscapes](https:\u002F\u002Fwww.cityscapes-dataset.com\u002F) | 512x256 | `fcn-resnet18-cityscapes-512x256` | 83.3% | 48 FPS | 480 FPS |\n| [Cityscapes](https:\u002F\u002Fwww.cityscapes-dataset.com\u002F) | 1024x512 | `fcn-resnet18-cityscapes-1024x512` | 87.3% | 12 FPS | 175 FPS |\n| [Cityscapes](https:\u002F\u002Fwww.cityscapes-dataset.com\u002F) | 2048x1024 | `fcn-resnet18-cityscapes-2048x1024` | 89.6% | 3 FPS | 47 FPS |\n| [DeepScene](http:\u002F\u002Fdeepscene.cs.uni-freiburg.de\u002F) | 576x320 | `fcn-resnet18-deepscene-576x320` | 96.4% | 26 FPS | 360 FPS |\n| [DeepScene](http:\u002F\u002Fdeepscene.cs.uni-freiburg.de\u002F) | 864x480 | `fcn-resnet18-deepscene-864x480` | 96.9% | 14 FPS | 190 FPS |\n| [Multi-Human](https:\u002F\u002Flv-mhp.github.io\u002F) | 512x320 | `fcn-resnet18-mhp-512x320` | 86.5% | 34 FPS | 370 FPS |\n| [Multi-Human](https:\u002F\u002Flv-mhp.github.io\u002F) | 640x360 | `fcn-resnet18-mhp-512x320` | 87.1% | 23 FPS | 325 FPS |\n| [Pascal VOC](http:\u002F\u002Fhost.robots.ox.ac.uk\u002Fpascal\u002FVOC\u002F) | 320x320 | `fcn-resnet18-voc-320x320` | 85.9% | 45 FPS | 508 FPS |\n| [Pascal VOC](http:\u002F\u002Fhost.robots.ox.ac.uk\u002Fpascal\u002FVOC\u002F) | 512x320 | `fcn-resnet18-voc-512x320` | 88.5% | 34 FPS | 375 FPS |\n| [SUN RGB-D](http:\u002F\u002Frgbd.cs.princeton.edu\u002F) | 512x400 | `fcn-resnet18-sun-512x400` | 64.3% | 28 FPS | 340 FPS |\n| [SUN RGB-D](http:\u002F\u002Frgbd.cs.princeton.edu\u002F) | 640x512 | `fcn-resnet18-sun-640x512` | 65.1% | 17 FPS | 224 FPS |\n\n* If the resolution is omitted from the CLI argument, the lowest resolution model is loaded\n* Accuracy indicates the pixel classification accuracy across the model's validation dataset\n* Performance is measured for GPU FP16 mode with JetPack 4.2.1, `nvpmodel 0` (MAX-N)\n\n\u003Cdetails>\n\u003Csummary>Legacy Segmentation Models\u003C\u002Fsummary>\n\n| Network                 | CLI Argument                    | NetworkType enum                | Classes |\n| ------------------------|---------------------------------|---------------------------------|---------|\n| Cityscapes (2048x2048)  | `fcn-alexnet-cityscapes-hd`     | `FCN_ALEXNET_CITYSCAPES_HD`     |    21   |\n| Cityscapes (1024x1024)  | `fcn-alexnet-cityscapes-sd`     | `FCN_ALEXNET_CITYSCAPES_SD`     |    21   |\n| Pascal VOC (500x356)    | `fcn-alexnet-pascal-voc`        | `FCN_ALEXNET_PASCAL_VOC`        |    21   |\n| Synthia (CVPR16)        | `fcn-alexnet-synthia-cvpr`      | `FCN_ALEXNET_SYNTHIA_CVPR`      |    14   |\n| Synthia (Summer-HD)     | `fcn-alexnet-synthia-summer-hd` | `FCN_ALEXNET_SYNTHIA_SUMMER_HD` |    14   |\n| Synthia (Summer-SD)     | `fcn-alexnet-synthia-summer-sd` | `FCN_ALEXNET_SYNTHIA_SUMMER_SD` |    14   |\n| Aerial-FPV (1280x720)   | `fcn-alexnet-aerial-fpv-720p`   | `FCN_ALEXNET_AERIAL_FPV_720p`   |     2   |\n\n\u003C\u002Fdetails>\n\n#### Pose Estimation\n\n| Model                   | CLI argument       | NetworkType enum   | Keypoints |\n| ------------------------|--------------------|--------------------|-----------|\n| Pose-ResNet18-Body      | `resnet18-body`    | `RESNET18_BODY`    | 18        |\n| Pose-ResNet18-Hand      | `resnet18-hand`    | `RESNET18_HAND`    | 21        |\n| Pose-DenseNet121-Body   | `densenet121-body` | `DENSENET121_BODY` | 18        |\n\n#### Action Recognition\n\n| Model                    | CLI argument | Classes |\n| -------------------------|--------------|---------|\n| Action-ResNet18-Kinetics | `resnet18`   |  1040   |\n| Action-ResNet34-Kinetics | `resnet34`   |  1040   |\n\n## Recommended System Requirements\n\n* Jetson Nano Developer Kit with JetPack 4.2 or newer (Ubuntu 18.04 aarch64).  \n* Jetson Nano 2GB Developer Kit with JetPack 4.4.1 or newer (Ubuntu 18.04 aarch64).\n* Jetson Orin Nano Developer Kit with JetPack 5.0 or newer (Ubuntu 20.04 aarch64).\n* Jetson Xavier NX Developer Kit with JetPack 4.4 or newer (Ubuntu 18.04 aarch64).  \n* Jetson AGX Xavier Developer Kit with JetPack 4.0 or newer (Ubuntu 18.04 aarch64).  \n* Jetson AGX Orin Developer Kit with JetPack 5.0 or newer (Ubuntu 20.04 aarch64).\n* Jetson TX2 Developer Kit with JetPack 3.0 or newer (Ubuntu 16.04 aarch64).  \n* Jetson TX1 Developer Kit with JetPack 2.3 or newer (Ubuntu 16.04 aarch64).  \n\nThe [Transfer Learning with PyTorch](#training) section of the tutorial speaks from the perspective of running PyTorch onboard Jetson for training DNNs, however the same PyTorch code can be used on a PC, server, or cloud instance with an NVIDIA discrete GPU for faster training.\n\n\n## Extra Resources\n\nIn this area, links and resources for deep learning are listed:\n\n* [ros_deep_learning](http:\u002F\u002Fwww.github.com\u002Fdusty-nv\u002Fros_deep_learning) - TensorRT inference ROS nodes\n* [NVIDIA AI IoT](https:\u002F\u002Fgithub.com\u002FNVIDIA-AI-IOT) - NVIDIA Jetson GitHub repositories\n* [Jetson eLinux Wiki](https:\u002F\u002Fwww.eLinux.org\u002FJetson) - Jetson eLinux Wiki\n\n\n## Two Days to a Demo (DIGITS)\n\n> **note:** the DIGITS\u002FCaffe tutorial from below is deprecated.  It's recommended to follow the [Transfer Learning with PyTorch](#training) tutorial from Hello AI World.\n \n\u003Cdetails>\n\u003Csummary>Expand this section to see original DIGITS tutorial (deprecated)\u003C\u002Fsummary>\n\u003Cbr\u002F>\nThe DIGITS tutorial includes training DNN's in the cloud or PC, and inference on the Jetson with TensorRT, and can take roughly two days or more depending on system setup, downloading the datasets, and the training speed of your GPU.\n\n* [DIGITS Workflow](docs\u002Fdigits-workflow.md) \n* [DIGITS System Setup](docs\u002Fdigits-setup.md)\n* [Setting up Jetson with JetPack](docs\u002Fjetpack-setup.md)\n* [Building the Project from Source](docs\u002Fbuilding-repo.md)\n* [Classifying Images with ImageNet](docs\u002Fimagenet-console.md)\n\t* [Using the Console Program on Jetson](docs\u002Fimagenet-console.md#using-the-console-program-on-jetson)\n\t* [Coding Your Own Image Recognition Program](docs\u002Fimagenet-example.md)\n\t* [Running the Live Camera Recognition Demo](docs\u002Fimagenet-camera.md)\n\t* [Re-Training the Network with DIGITS](docs\u002Fimagenet-training.md)\n\t* [Downloading Image Recognition Dataset](docs\u002Fimagenet-training.md#downloading-image-recognition-dataset)\n\t* [Customizing the Object Classes](docs\u002Fimagenet-training.md#customizing-the-object-classes)\n\t* [Importing Classification Dataset into DIGITS](docs\u002Fimagenet-training.md#importing-classification-dataset-into-digits)\n\t* [Creating Image Classification Model with DIGITS](docs\u002Fimagenet-training.md#creating-image-classification-model-with-digits)\n\t* [Testing Classification Model in DIGITS](docs\u002Fimagenet-training.md#testing-classification-model-in-digits)\n\t* [Downloading Model Snapshot to Jetson](docs\u002Fimagenet-snapshot.md)\n\t* [Loading Custom Models on Jetson](docs\u002Fimagenet-custom.md)\n* [Locating Objects with DetectNet](docs\u002Fdetectnet-training.md)\n\t* [Detection Data Formatting in DIGITS](docs\u002Fdetectnet-training.md#detection-data-formatting-in-digits)\n\t* [Downloading the Detection Dataset](docs\u002Fdetectnet-training.md#downloading-the-detection-dataset)\n\t* [Importing the Detection Dataset into DIGITS](docs\u002Fdetectnet-training.md#importing-the-detection-dataset-into-digits)\n\t* [Creating DetectNet Model with DIGITS](docs\u002Fdetectnet-training.md#creating-detectnet-model-with-digits)\n\t* [Testing DetectNet Model Inference in DIGITS](docs\u002Fdetectnet-training.md#testing-detectnet-model-inference-in-digits)\n\t* [Downloading the Detection Model to Jetson](docs\u002Fdetectnet-snapshot.md)\n\t* [DetectNet Patches for TensorRT](docs\u002Fdetectnet-snapshot.md#detectnet-patches-for-tensorrt)\n\t* [Detecting Objects from the Command Line](docs\u002Fdetectnet-console.md)\n\t* [Multi-class Object Detection Models](docs\u002Fdetectnet-console.md#multi-class-object-detection-models)\n\t* [Running the Live Camera Detection Demo on Jetson](docs\u002Fdetectnet-camera.md)\n* [Semantic Segmentation with SegNet](docs\u002Fsegnet-dataset.md)\n\t* [Downloading Aerial Drone Dataset](docs\u002Fsegnet-dataset.md#downloading-aerial-drone-dataset)\n\t* [Importing the Aerial Dataset into DIGITS](docs\u002Fsegnet-dataset.md#importing-the-aerial-dataset-into-digits)\n\t* [Generating Pretrained FCN-Alexnet](docs\u002Fsegnet-pretrained.md)\n\t* [Training FCN-Alexnet with DIGITS](docs\u002Fsegnet-training.md)\n\t* [Testing Inference Model in DIGITS](docs\u002Fsegnet-training.md#testing-inference-model-in-digits)\n\t* [FCN-Alexnet Patches for TensorRT](docs\u002Fsegnet-patches.md)\n\t* [Running Segmentation Models on Jetson](docs\u002Fsegnet-console.md)\n\n\u003C\u002Fdetails>\n\n##\n\u003Cp align=\"center\">\u003Csup>© 2016-2019 NVIDIA | \u003C\u002Fsup>\u003Ca href=\"#deploying-deep-learning\">\u003Csup>Table of Contents\u003C\u002Fsup>\u003C\u002Fa>\u003C\u002Fp>\n\n","\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_a68fe911e22b.jpg\" width=\"100%\">\n\n# 部署深度学习\n欢迎来到我们针对 **[NVIDIA Jetson](https:\u002F\u002Fdeveloper.nvidia.com\u002Fembedded-computing)** 设备的推理与实时视觉 [DNN 库](#api-reference) 教程。该项目使用 **[TensorRT](https:\u002F\u002Fdeveloper.nvidia.com\u002Ftensorrt)** 在 GPU 上以 C++ 或 Python 运行优化后的网络，并利用 PyTorch 进行模型训练。\n\n支持的 DNN 视觉原语包括用于图像分类的 [`imageNet`](docs\u002Fimagenet-console-2.md)、用于目标检测的 [`detectNet`](docs\u002Fdetectnet-console-2.md)、用于语义分割的 [`segNet`](docs\u002Fsegnet-console-2.md)、用于姿态估计的 [`poseNet`](docs\u002Fposenet.md)，以及用于动作识别的 [`actionNet`](docs\u002Factionnet.md)。我们还提供了从实时摄像头流中进行处理、使用 WebRTC 构建 Web 应用，以及对 ROS\u002FROS2 的支持等示例。\n\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_8b9f587bb1c5.jpg\">\n\n请按照 **[Hello AI World](#hello-ai-world)** 教程，在您的 Jetson 设备上运行推理和迁移学习，其中包括收集您自己的数据集、使用 PyTorch 训练自定义模型，并通过 TensorRT 部署这些模型。\n\n### 目录\n\n* [Hello AI World](#hello-ai-world)\n* [Jetson AI Lab](#jetson-ai-lab)\n* [视频教程](#video-walkthroughs)\n* [API 参考](#api-reference)\n* [代码示例](#code-examples)\n* [预训练模型](#pre-trained-models)\n* [系统要求](#recommended-system-requirements)\n* [变更日志](CHANGELOG.md)\n\n> >   JetPack 6 现已支持 Orin 设备 ([developer.nvidia.com\u002Fjetpack](https:\u002F\u002Fdeveloper.nvidia.com\u002Fembedded\u002Fjetpack)) \u003Cbr\u002F>\n> >   欢迎访问 [Jetson AI Lab](https:\u002F\u002Fwww.jetson-ai-lab.com\u002F) 查看生成式 AI 和 LLM 教程！ \u003Cbr\u002F>\n> >   请参阅 [变更日志](CHANGELOG.md) 以获取最新更新和新功能。 \u003Cbr\u002F>\n\n## Hello AI World\nHello AI World 可以完全在您的 Jetson 设备上运行，包括使用 TensorRT 进行实时推理，以及利用 PyTorch 进行迁移学习。有关安装说明，请参阅 [系统设置](#system-setup)。建议先从 [推理](#inference) 部分开始，熟悉相关概念，然后再深入学习 [训练](#training) 自己的模型。\n\n#### 系统设置\n\n* [使用 JetPack 设置 Jetson](docs\u002Fjetpack-setup-2.md)\n* [运行 Docker 容器](docs\u002Faux-docker.md)\n* [从源码构建项目](docs\u002Fbuilding-repo-2.md)\n\n#### 推理\n\n* [图像分类](docs\u002Fimagenet-console-2.md)\n\t* [在 Jetson 上使用 ImageNet 程序](docs\u002Fimagenet-console-2.md)\n\t* [编写您自己的图像识别程序（Python）](docs\u002Fimagenet-example-python-2.md)\n\t* [编写您自己的图像识别程序（C++）](docs\u002Fimagenet-example-2.md)\n\t* [运行实时摄像头识别演示](docs\u002Fimagenet-camera-2.md)\n\t* [用于图像标签的多标签分类](docs\u002Fimagenet-tagging.md)\n* [目标检测](docs\u002Fdetectnet-console-2.md)\n\t* [从图像中检测目标](docs\u002Fdetectnet-console-2.md#detecting-objects-from-the-command-line)\n\t* [运行实时摄像头检测演示](docs\u002Fdetectnet-camera-2.md)\n\t* [编写您自己的目标检测程序](docs\u002Fdetectnet-example-2.md)\n\t* [使用 TAO 检测模型](docs\u002Fdetectnet-tao.md)\n\t* [视频中的目标跟踪](docs\u002Fdetectnet-tracking.md)\n* [语义分割](docs\u002Fsegnet-console-2.md)\n\t* [从命令行分割图像](docs\u002Fsegnet-console-2.md#segmenting-images-from-the-command-line)\n\t* [运行实时摄像头分割演示](docs\u002Fsegnet-camera-2.md)\n* [姿态估计](docs\u002Fposenet.md)\n* [动作识别](docs\u002Factionnet.md)\n* [背景去除](docs\u002Fbackgroundnet.md)\n* [单目深度](docs\u002Fdepthnet.md)\n\n#### 训练\n\n* [使用 PyTorch 进行迁移学习](docs\u002Fpytorch-transfer-learning.md)\n* 分类\u002F识别（ResNet-18）\n\t* [基于猫狗数据集的再训练](docs\u002Fpytorch-cat-dog.md)\n\t* [基于 PlantCLEF 数据集的再训练](docs\u002Fpytorch-plants.md)\n\t* [收集您自己的分类数据集](docs\u002Fpytorch-collect.md)\n* 目标检测（SSD-Mobilenet）\n\t* [再训练 SSD-Mobilenet](docs\u002Fpytorch-ssd.md)\n\t* [收集您自己的检测数据集](docs\u002Fpytorch-collect-detection.md)\n\n#### Web 应用框架\n\n* [WebRTC 服务器](docs\u002Fwebrtc-server.md)\n* [HTML \u002F JavaScript](docs\u002Fwebrtc-html.md)\n* [Flask + REST](docs\u002Fwebrtc-flask.md)\n* [Plotly 仪表板](docs\u002Fwebrtc-dash.md)\n* [Recognizer（交互式训练）](docs\u002Fwebrtc-recognizer.md)\n\n#### 附录\n\n* [摄像头流媒体与多媒体](docs\u002Faux-streaming.md)\n* [使用 CUDA 进行图像处理](docs\u002Faux-image.md)\n* [适用于 ROS\u002FROS2 的 DNN 推理节点](https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fros_deep_learning)\n\n## Jetson AI Lab\n\u003Ca href=\"https:\u002F\u002Fwww.jetson-ai-lab.com\">\u003Cimg align=\"right\" width=\"200\" height=\"200\" src=\"https:\u002F\u002Fnvidia-ai-iot.github.io\u002Fjetson-generative-ai-playground\u002Fimages\u002FJON_Gen-AI-panels.png\">\u003C\u002Fa>\n\n[**Jetson AI Lab**](https:\u002F\u002Fwww.jetson-ai-lab.com) 提供了更多关于 LLM、视觉 Transformer (ViT) 和视觉语言模型 (VLM) 的教程，这些模型可在 Orin（有时也可在 Xavier 上）运行。以下是一些示例：\n\n\u003Ca href=\"https:\u002F\u002Fwww.jetson-ai-lab.com\u002Ftutorial_nanoowl.html\">\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_5dcd1c4bb5c7.gif\">\u003C\u002Fa>\n> [NanoOWL - 开放词汇目标检测 ViT](https:\u002F\u002Fwww.jetson-ai-lab.com\u002Ftutorial_nanoowl.html)（容器：[`nanoowl`](\u002Fpackages\u002Fvit\u002Fnanoowl)）\n\n\u003Ca href=\"https:\u002F\u002Fyoutu.be\u002FX-OXxPiUTuU\">\u003Cimg width=\"600px\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_63f2240c52ae.gif\">\u003C\u002Fa>\n> [Jetson AGX Orin 上的 Live Llava](https:\u002F\u002Fyoutu.be\u002FX-OXxPiUTuU)（容器：[`local_llm`](\u002Fpackages\u002Fllm\u002Flocal_llm#live-llava)）\n\n\u003Ca href=\"https:\u002F\u002Fyoutu.be\u002FdRmAGGuupuE\">\u003Cimg width=\"600px\" src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_04332a9ffc81.jpg\">\u003C\u002Fa>\n> [Live Llava 2.0 - VILA + 多模态 NanoDB 在 Jetson Orin 上](https:\u002F\u002Fyoutu.be\u002FX-OXxPiUTuU)（容器：[`local_llm`](\u002Fpackages\u002Fllm\u002Flocal_llm#live-llava)）\n\n\u003Ca href=\"https:\u002F\u002Fyoutu.be\u002FayqKpQNd1Jw\">\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_readme_caf4cb22628b.gif\">\u003C\u002Fa>\n> [NVIDIA Jetson 上的实时多模态向量数据库](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=wzLHAgDxMjQ)（容器：[`nanodb`](\u002Fpackages\u002Fvectordb\u002Fnanodb)）\n\n## 视频教程\n\n以下是为 [Jetson AI 认证](https:\u002F\u002Fdeveloper.nvidia.com\u002Fembedded\u002Flearn\u002Fjetson-ai-certification-programs) 课程录制的 Hello AI World 屏幕录像：\n\n| 描述                                                                                                                                                                                                                                                                                                        | 视频                                                                                                                                                                                                                                                 |\n|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QXIwdsyK7Rw&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=9\" target=\"_blank\">**Hello AI World 设置**\u003C\u002Fa>\u003Cbr\u002F>在 Jetson Nano 上下载并运行 Hello AI World 容器，测试摄像头画面，并了解如何通过 RTP 协议将其在网络中进行流式传输。                                     | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QXIwdsyK7Rw&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=9\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_setup.jpg width=\"750\">\u003C\u002Fa>               |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QatH8iF0Efk&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=10\" target=\"_blank\">**图像分类推理**\u003C\u002Fa>\u003Cbr\u002F>使用 Jetson Nano 和深度学习编写自己的 Python 程序进行图像分类，然后在实时摄像头流上尝试进行实时分类。 | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QatH8iF0Efk&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=10\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_imagenet.jpg width=\"750\">\u003C\u002Fa>           |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=sN6aT9TpltU&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=11\" target=\"_blank\">**训练图像分类模型**\u003C\u002Fa>\u003Cbr\u002F>学习如何在 Jetson Nano 上使用 PyTorch 训练图像分类模型，并收集自己的分类数据集以创建自定义模型。     | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=sN6aT9TpltU&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=11\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_imagenet_training.jpg width=\"750\">\u003C\u002Fa>  |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=obt60r8ZeB0&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=12\" target=\"_blank\">**目标检测推理**\u003C\u002Fa>\u003Cbr\u002F>使用 Jetson Nano 和深度学习编写自己的 Python 程序进行目标检测，然后在实时摄像头流上尝试进行实时检测。              | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=obt60r8ZeB0&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=12\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_detectnet.jpg width=\"750\">\u003C\u002Fa>          |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=2XMkPW_sIGg&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=13\" target=\"_blank\">**训练目标检测模型**\u003C\u002Fa>\u003Cbr\u002F>学习如何在 Jetson Nano 上使用 PyTorch 训练目标检测模型，并收集自己的检测数据集以创建自定义模型。                  | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=2XMkPW_sIGg&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=13\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_detectnet_training.jpg width=\"750\">\u003C\u002Fa> |\n| \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=AQhkMLaB_fY&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=14\" target=\"_blank\">**语义分割**\u003C\u002Fa>\u003Cbr\u002F>在 Jetson Nano 上尝试全卷积语义分割网络，并在实时摄像头流上运行实时分割。                                 | \u003Ca href=\"https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=AQhkMLaB_fY&list=PL5B692fm6--uQRRDTPsJDp4o0xbzkoyf8&index=14\" target=\"_blank\">\u003Cimg src=https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fraw\u002Fmaster\u002Fdocs\u002Fimages\u002Fthumbnail_segnet.jpg width=\"750\">\u003C\u002Fa>             |\n\n## API 参考\n\n以下是来自该仓库的 [C++](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Findex.html) 和 [Python](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.html) 库的参考文档链接：\n\n#### jetson-inference\n\n|                    | [C++](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__deepVision.html) | [Python](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html) |\n|--------------------|--------------|--------------|\n| 图像识别  | [`imageNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__imageNet.html#classimageNet) | [`imageNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#imageNet) |\n| 目标检测   | [`detectNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__detectNet.html#classdetectNet) | [`detectNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#detectNet)\n| 分割       | [`segNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__segNet.html#classsegNet) | [`segNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#segNet) |\n| 姿态估计    | [`poseNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__poseNet.html#classposeNet) | [`poseNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#poseNet) |\n| 行为识别 | [`actionNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__actionNet.html#classactionNet) | [`actionNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#actionNet) |\n| 背景去除 | [`backgroundNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__backgroundNet.html#classbackgroundNet) | [`actionNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#backgroundNet) |\n| 单目深度    | [`depthNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__depthNet.html#classdepthNet) | [`depthNet`](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.inference.html#depthNet) |\n\n#### jetson-utils\n\n* [C++](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fgroup__util.html)\n* [Python](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Fpython\u002Fjetson.utils.html)\n\n这些库可以通过链接到 `libjetson-inference` 和 `libjetson-utils` 来在外部项目中使用。\n\n## 代码示例\n\nHello AI World 教程中的以下步骤涵盖了使用该库的入门级代码讲解：\n\n* [编写你自己的图像识别程序（Python）](docs\u002Fimagenet-example-python-2.md)\n* [编写你自己的图像识别程序（C++）](docs\u002Fimagenet-example-2.md)\n\n此外，还有更多用于在图像和实时摄像头流上运行神经网络的 C++ 和 Python 示例，如下所示：\n\n|                   | C++              | Python             |\n|-------------------|---------------------|---------------------|\n|    图像识别  | [`imagenet.cpp`](examples\u002Fimagenet\u002Fimagenet.cpp) | [`imagenet.py`](python\u002Fexamples\u002Fimagenet.py) |\n|    目标检测   | [`detectnet.cpp`](examples\u002Fdetectnet\u002Fdetectnet.cpp) | [`detectnet.py`](python\u002Fexamples\u002Fdetectnet.py) |\n|    分割       | [`segnet.cpp`](examples\u002Fsegnet\u002Fsegnet.cpp) | [`segnet.py`](python\u002Fexamples\u002Fsegnet.py) |\n|    姿态估计    | [`posenet.cpp`](examples\u002Fposenet\u002Fposenet.cpp) | [`posenet.py`](python\u002Fexamples\u002Fposenet.py) |\n|    行为识别 | [`actionnet.cpp`](examples\u002Factionnet\u002Factionnet.cpp) | [`actionnet.py`](python\u002Fexamples\u002Factionnet.py) |\n|    背景去除 | [`backgroundnet.cpp`](examples\u002Fbackgroundnet\u002Fbackgroundnet.cpp) | [`backgroundnet.py`](python\u002Fexamples\u002Fbackgroundnet.py) |\n|    单目深度    | [`depthnet.cpp`](examples\u002Fdepthnet\u002Fsegnet.cpp) | [`depthnet.py`](python\u002Fexamples\u002Fdepthnet.py) |\n\n> **注**：有关如何与其他 Python 库（如 Numpy、PyTorch 等）一起使用内存，请参阅 [数组接口](docs\u002Faux-image.md#array-interfaces) 部分。\n\n这些示例将在[从源代码构建项目](docs\u002Fbuilding-repo-2.md)时自动编译，并且除了用户提供的自定义模型外，还可以运行下面列出的预训练模型。每个示例都可以通过添加 `--help` 参数来查看使用说明。\n\n## 预训练模型\n\n该项目附带多个预训练模型，可供直接使用，并会自动下载：\n\n#### 图像识别\n\n| 网络       | CLI 参数   | NetworkType 枚举 |\n| --------------|----------------|------------------|\n| AlexNet       | `alexnet`      | `ALEXNET`        |\n| GoogleNet     | `googlenet`    | `GOOGLENET`      |\n| GoogleNet-12  | `googlenet-12` | `GOOGLENET_12`   |\n| ResNet-18     | `resnet-18`    | `RESNET_18`      |\n| ResNet-50     | `resnet-50`    | `RESNET_50`      |\n| ResNet-101    | `resnet-101`   | `RESNET_101`     |\n| ResNet-152    | `resnet-152`   | `RESNET_152`     |\n| VGG-16        | `vgg-16`       | `VGG-16`         |\n| VGG-19        | `vgg-19`       | `VGG-19`         |\n| Inception-v4  | `inception-v4` | `INCEPTION_V4`   |\n\n#### 目标检测\n\n| 模型                   | CLI 参数       | NetworkType 枚举   | 物体类别       |\n| ------------------------|--------------------|--------------------|----------------------|\n| SSD-Mobilenet-v1        | `ssd-mobilenet-v1` | `SSD_MOBILENET_V1` | 91 个（[COCO 类别](..\u002Fdata\u002Fnetworks\u002Fssd_coco_labels.txt)）     |\n| SSD-Mobilenet-v2        | `ssd-mobilenet-v2` | `SSD_MOBILENET_V2` | 91 个（[COCO 类别](..\u002Fdata\u002Fnetworks\u002Fssd_coco_labels.txt)）     |\n| SSD-Inception-v2        | `ssd-inception-v2` | `SSD_INCEPTION_V2` | 91 个（[COCO 类别](..\u002Fdata\u002Fnetworks\u002Fssd_coco_labels.txt)）     |\n| TAO PeopleNet           | `peoplenet`        | `PEOPLENET`        | 人、包、脸    |\n| TAO PeopleNet（剪枝版）  | `peoplenet-pruned` | `PEOPLENET_PRUNED` | 人、包、脸    |\n| TAO DashCamNet          | `dashcamnet`       | `DASHCAMNET`       | 人、车、自行车、交通标志 |\n| TAO TrafficCamNet       | `trafficcamnet`    | `TRAFFICCAMNET`    | 人、车、自行车、交通标志 | \n| TAO FaceDetect          | `facedetect`       | `FACEDETECT`       | 脸                 |\n\n\u003Cdetails>\n\u003Csummary>旧版检测模型\u003C\u002Fsummary>\n\n| 模型                   | CLI 参数       | NetworkType 枚举   | 物体类别       |\n| ------------------------|--------------------|--------------------|----------------------|\n| DetectNet-COCO-Dog      | `coco-dog`         | `COCO_DOG`         | 狗                 |\n| DetectNet-COCO-Bottle   | `coco-bottle`      | `COCO_BOTTLE`      | 瓶子              |\n| DetectNet-COCO-Chair    | `coco-chair`       | `COCO_CHAIR`       | 椅子               |\n| DetectNet-COCO-Airplane | `coco-airplane`    | `COCO_AIRPLANE`    | 飞机            |\n| ped-100                 | `pednet`           | `PEDNET`           | 行人          |\n| multiped-500            | `multiped`         | `PEDNET_MULTI`     | 行人、行李 |\n| facenet-120             | `facenet`          | `FACENET`          | 面部                |\n\n\u003C\u002Fdetails>\n\n#### 语义分割\n\n| 数据集      | 分辨率 | CLI 参数 | 精度 | Jetson Nano | Jetson Xavier |\n|:------------:|:----------:|--------------|:--------:|:-----------:|:-------------:|\n| [Cityscapes](https:\u002F\u002Fwww.cityscapes-dataset.com\u002F) | 512x256 | `fcn-resnet18-cityscapes-512x256` | 83.3% | 48 FPS | 480 FPS |\n| [Cityscapes](https:\u002F\u002Fwww.cityscapes-dataset.com\u002F) | 1024x512 | `fcn-resnet18-cityscapes-1024x512` | 87.3% | 12 FPS | 175 FPS |\n| [Cityscapes](https:\u002F\u002Fwww.cityscapes-dataset.com\u002F) | 2048x1024 | `fcn-resnet18-cityscapes-2048x1024` | 89.6% | 3 FPS | 47 FPS |\n| [DeepScene](http:\u002F\u002Fdeepscene.cs.uni-freiburg.de\u002F) | 576x320 | `fcn-resnet18-deepscene-576x320` | 96.4% | 26 FPS | 360 FPS |\n| [DeepScene](http:\u002F\u002Fdeepscene.cs.uni-freiburg.de\u002F) | 864x480 | `fcn-resnet18-deepscene-864x480` | 96.9% | 14 FPS | 190 FPS |\n| [Multi-Human](https:\u002F\u002Flv-mhp.github.io\u002F) | 512x320 | `fcn-resnet18-mhp-512x320` | 86.5% | 34 FPS | 370 FPS |\n| [Multi-Human](https:\u002F\u002Flv-mhp.github.io\u002F) | 640x360 | `fcn-resnet18-mhp-512x320` | 87.1% | 23 FPS | 325 FPS |\n| [Pascal VOC](http:\u002F\u002Fhost.robots.ox.ac.uk\u002Fpascal\u002FVOC\u002F) | 320x320 | `fcn-resnet18-voc-320x320` | 85.9% | 45 FPS | 508 FPS |\n| [Pascal VOC](http:\u002F\u002Fhost.robots.ox.ac.uk\u002Fpascal\u002FVOC\u002F) | 512x320 | `fcn-resnet18-voc-512x320` | 88.5% | 34 FPS | 375 FPS |\n| [SUN RGB-D](http:\u002F\u002Frgbd.cs.princeton.edu\u002F) | 512x400 | `fcn-resnet18-sun-512x400` | 64.3% | 28 FPS | 340 FPS |\n| [SUN RGB-D](http:\u002F\u002Frgbd.cs.princeton.edu\u002F) | 640x512 | `fcn-resnet18-sun-640x512` | 65.1% | 17 FPS | 224 FPS |\n\n* 如果 CLI 参数中未指定分辨率，则加载最低分辨率的模型\n* 精度表示模型在验证数据集上的像素分类准确率\n* 性能是在 JetPack 4.2.1、`nvpmodel 0`（MAX-N）设置下，以 GPU FP16 模式测量的\n\n\u003Cdetails>\n\u003Csummary>旧版分割模型\u003C\u002Fsummary>\n\n| 网络                 | CLI 参数                    | NetworkType 枚举                | 类别 |\n| ------------------------|---------------------------------|---------------------------------|---------|\n| Cityscapes (2048x2048)  | `fcn-alexnet-cityscapes-hd`     | `FCN_ALEXNET_CITYSCAPES_HD`     |    21   |\n| Cityscapes (1024x1024)  | `fcn-alexnet-cityscapes-sd`     | `FCN_ALEXNET_CITYSCAPES_SD`     |    21   |\n| Pascal VOC (500x356)    | `fcn-alexnet-pascal-voc`        | `FCN_ALEXNET_PASCAL_VOC`        |    21   |\n| Synthia (CVPR16)        | `fcn-alexnet-synthia-cvpr`      | `FCN_ALEXNET_SYNTHIA_CVPR`      |    14   |\n| Synthia (Summer-HD)     | `fcn-alexnet-synthia-summer-hd` | `FCN_ALEXNET_SYNTHIA_SUMMER_HD` |    14   |\n| Synthia (Summer-SD)     | `fcn-alexnet-synthia-summer-sd` | `FCN_ALEXNET_SYNTHIA_SUMMER_SD` |    14   |\n| Aerial-FPV (1280x720)   | `fcn-alexnet-aerial-fpv-720p`   | `FCN_ALEXNET_AERIAL_FPV_720p`   |     2   |\n\n\u003C\u002Fdetails>\n\n#### 姿态估计\n\n| 模型                   | CLI 参数       | NetworkType 枚举   | 关键点 |\n| ------------------------|--------------------|--------------------|-----------|\n| Pose-ResNet18-Body      | `resnet18-body`    | `RESNET18_BODY`    | 18        |\n| Pose-ResNet18-Hand      | `resnet18-hand`    | `RESNET18_HAND`    | 21        |\n| Pose-DenseNet121-Body   | `densenet121-body` | `DENSENET121_BODY` | 18        |\n\n#### 动作识别\n\n| 模型                    | CLI 参数 | 类别 |\n| -------------------------|--------------|---------|\n| Action-ResNet18-Kinetics | `resnet18`   |  1040   |\n| Action-ResNet34-Kinetics | `resnet34`   |  1040   |\n\n## 推荐系统要求\n\n* Jetson Nano 开发者套件，配备 JetPack 4.2 或更高版本（Ubuntu 18.04 aarch64）。  \n* Jetson Nano 2GB 开发者套件，配备 JetPack 4.4.1 或更高版本（Ubuntu 18.04 aarch64）。\n* Jetson Orin Nano 开发者套件，配备 JetPack 5.0 或更高版本（Ubuntu 20.04 aarch64）。\n* Jetson Xavier NX 开发者套件，配备 JetPack 4.4 或更高版本（Ubuntu 18.04 aarch64）。  \n* Jetson AGX Xavier 开发者套件，配备 JetPack 4.0 或更高版本（Ubuntu 18.04 aarch64）。  \n* Jetson AGX Orin 开发者套件，配备 JetPack 5.0 或更高版本（Ubuntu 20.04 aarch64）。\n* Jetson TX2 开发者套件，配备 JetPack 3.0 或更高版本（Ubuntu 16.04 aarch64）。  \n* Jetson TX1 开发者套件，配备 JetPack 2.3 或更高版本（Ubuntu 16.04 aarch64）。  \n\n教程中的[使用 PyTorch 进行迁移学习](#training)部分是从在 Jetson 上运行 PyTorch 来训练深度神经网络的角度出发的，然而相同的 PyTorch 代码也可以在配备 NVIDIA 独立显卡的 PC、服务器或云实例上使用，以实现更快的训练。\n\n\n## 额外资源\n\n在此部分列出了深度学习相关的链接和资源：\n\n* [ros_deep_learning](http:\u002F\u002Fwww.github.com\u002Fdusty-nv\u002Fros_deep_learning) - TensorRT 推理 ROS 节点\n* [NVIDIA AI IoT](https:\u002F\u002Fgithub.com\u002FNVIDIA-AI-IOT) - NVIDIA Jetson GitHub 仓库\n* [Jetson eLinux Wiki](https:\u002F\u002Fwww.eLinux.org\u002FJetson) - Jetson eLinux 维基百科\n\n\n## 两天打造演示（DIGITS）\n\n> **注意：** 下文中的 DIGITS\u002FCaffe 教程已弃用。建议您参考 Hello AI World 中的[使用 PyTorch 进行迁移学习](#training)教程。\n\n\u003Cdetails>\n\u003Csummary>展开此部分以查看原始的 DIGITS 教程（已弃用）\u003C\u002Fsummary>\n\u003Cbr\u002F>\nDIGITS 教程包括在云端或 PC 上训练深度神经网络，以及在 Jetson 上使用 TensorRT 进行推理，整个过程大约需要两天或更长时间，具体取决于系统配置、数据集的下载以及 GPU 的训练速度。\n\n* [DIGITS 工作流程](docs\u002Fdigits-workflow.md) \n* [DIGITS 系统设置](docs\u002Fdigits-setup.md)\n* [在 Jetson 上安装 JetPack](docs\u002Fjetpack-setup.md)\n* [从源代码构建项目](docs\u002Fbuilding-repo.md)\n* [使用 ImageNet 对图像进行分类](docs\u002Fimagenet-console.md)\n\t* [在 Jetson 上使用控制台程序](docs\u002Fimagenet-console.md#using-the-console-program-on-jetson)\n\t* [编写自己的图像识别程序](docs\u002Fimagenet-example.md)\n\t* [运行实时摄像头识别演示](docs\u002Fimagenet-camera.md)\n\t* [使用 DIGITS 重新训练网络](docs\u002Fimagenet-training.md)\n\t* [下载图像识别数据集](docs\u002Fimagenet-training.md#downloading-image-recognition-dataset)\n\t* [自定义目标类别](docs\u002Fimagenet-training.md#customizing-the-object-classes)\n\t* [将分类数据集导入 DIGITS](docs\u002Fimagenet-training.md#importing-classification-dataset-into-digits)\n\t* [使用 DIGITS 创建图像分类模型](docs\u002Fimagenet-training.md#creating-image-classification-model-with-digits)\n\t* [在 DIGITS 中测试分类模型](docs\u002Fimagenet-training.md#testing-classification-model-in-digits)\n\t* [将模型快照下载到 Jetson](docs\u002Fimagenet-snapshot.md)\n\t* [在 Jetson 上加载自定义模型](docs\u002Fimagenet-custom.md)\n* [使用 DetectNet 定位物体](docs\u002Fdetectnet-training.md)\n\t* [在 DIGITS 中格式化检测数据](docs\u002Fdetectnet-training.md#detection-data-formatting-in-digits)\n\t* [下载检测数据集](docs\u002Fdetectnet-training.md#downloading-the-detection-dataset)\n\t* [将检测数据集导入 DIGITS](docs\u002Fdetectnet-training.md#importing-the-detection-dataset-into-digits)\n\t* [使用 DIGITS 创建 DetectNet 模型](docs\u002Fdetectnet-training.md#creating-detectnet-model-with-digits)\n\t* [在 DIGITS 中测试 DetectNet 模型的推理能力](docs\u002Fdetectnet-training.md#testing-detectnet-model-inference-in-digits)\n\t* [将检测模型下载到 Jetson](docs\u002Fdetectnet-snapshot.md)\n\t* [DetectNet 的 TensorRT 补丁](docs\u002Fdetectnet-snapshot.md#detectnet-patches-for-tensorrt)\n\t* [通过命令行检测物体](docs\u002Fdetectnet-console.md)\n\t* [多类目标检测模型](docs\u002Fdetectnet-console.md#multi-class-object-detection-models)\n\t* [在 Jetson 上运行实时摄像头检测演示](docs\u002Fdetectnet-camera.md)\n* [使用 SegNet 进行语义分割](docs\u002Fsegnet-dataset.md)\n\t* [下载航拍无人机数据集](docs\u002Fsegnet-dataset.md#downloading-aerial-drone-dataset)\n\t* [将航拍数据集导入 DIGITS](docs\u002Fsegnet-dataset.md#importing-the-aerial-dataset-into-digits)\n\t* [生成预训练的 FCN-Alexnet](docs\u002Fsegnet-pretrained.md)\n\t* [使用 DIGITS 训练 FCN-Alexnet](docs\u002Fsegnet-training.md)\n\t* [在 DIGITS 中测试推理模型](docs\u002Fsegnet-training.md#testing-inference-model-in-digits)\n\t* [FCN-Alexnet 的 TensorRT 补丁](docs\u002Fsegnet-patches.md)\n\t* [在 Jetson 上运行分割模型](docs\u002Fsegnet-console.md)\n\n\u003C\u002Fdetails>\n\n##\n\u003Cp align=\"center\">\u003Csup>© 2016-2019 NVIDIA | \u003C\u002Fsup>\u003Ca href=\"#deploying-deep-learning\">\u003Csup>目录\u003C\u002Fsup>\u003C\u002Fa>\u003C\u002Fp>","# jetson-inference 快速上手指南\n\n`jetson-inference` 是专为 **NVIDIA Jetson** 系列嵌入式设备设计的深度学习推理库。它基于 **TensorRT** 优化，支持通过 C++ 或 Python 调用预训练模型，实现图像分类、物体检测、语义分割、姿态估计等实时视觉任务，并支持使用 PyTorch 进行迁移学习训练。\n\n## 1. 环境准备\n\n### 系统要求\n*   **硬件平台**：NVIDIA Jetson 系列 (Nano, TX2, Xavier NX\u002FAGX, Orin Nano\u002FNX\u002FAGX)。\n*   **操作系统**：NVIDIA JetPack SDK (推荐 JetPack 5 或 6，Orin 系列需 JetPack 6)。\n*   **软件依赖**：\n    *   CUDA & cuDNN (随 JetPack 安装)\n    *   TensorRT (随 JetPack 安装)\n    *   PyTorch (用于训练)\n    *   Docker (推荐用于简化环境配置)\n\n> **注意**：本指南推荐使用官方提供的 Docker 容器运行，以避免复杂的本地依赖冲突。\n\n## 2. 安装步骤\n\n最快捷的方式是拉取预构建的 Docker 镜像并运行容器。\n\n### 步骤一：拉取 Docker 镜像\n根据你的 Jetson 架构选择对应的镜像（以下以通用指令为例，具体架构标签请参考 NVIDIA NGC 或项目文档）：\n\n```bash\ndocker pull dustynv\u002Fjetson-inference:r36.2.0\n# 注：r36.2.0 对应 JetPack 6 (Orin), 其他版本请查阅最新 tag\n```\n\n*国内开发者提示：如果拉取速度慢，可配置 Docker 国内镜像加速器（如阿里云、腾讯云等）。*\n\n### 步骤二：运行容器\n运行容器时需映射摄像头设备、显示服务和当前目录，以便访问本地文件和摄像头：\n\n```bash\ndocker run --runtime nvidia -it --rm --net=host \\\n  -e DISPLAY=$DISPLAY \\\n  -v \u002Ftmp\u002F.X11-unix:\u002Ftmp\u002F.X11-unix \\\n  -v $(pwd):\u002Fworkspace \\\n  --device \u002Fdev\u002Fvideo0:\u002Fdev\u002Fvideo0 \\\n  --device \u002Fdev\u002Fvideo1:\u002Fdev\u002Fvideo1 \\\n  --device \u002Fdev\u002Fvideo2:\u002Fdev\u002Fvideo2 \\\n  --device \u002Fdev\u002Fvideo3:\u002Fdev\u002Fvideo3 \\\n  --device \u002Fdev\u002Fvideo4:\u002Fdev\u002Fvideo4 \\\n  --device \u002Fdev\u002Fvideo5:\u002Fdev\u002Fvideo5 \\\n  --device \u002Fdev\u002Fvideo6:\u002Fdev\u002Fvideo6 \\\n  --device \u002Fdev\u002Fvideo7:\u002Fdev\u002Fvideo7 \\\n  --device \u002Fdev\u002Fvideo8:\u002Fdev\u002Fvideo8 \\\n  --device \u002Fdev\u002Fvideo9:\u002Fdev\u002Fvideo9 \\\n  --device \u002Fdev\u002Fvideo10:\u002Fdev\u002Fvideo10 \\\n  --device \u002Fdev\u002Fvideo11:\u002Fdev\u002Fvideo11 \\\n  --device \u002Fdev\u002Fvideo12:\u002Fdev\u002Fvideo12 \\\n  --device \u002Fdev\u002Fvideo13:\u002Fdev\u002Fvideo13 \\\n  --device \u002Fdev\u002Fvideo14:\u002Fdev\u002Fvideo14 \\\n  --device \u002Fdev\u002Fvideo15:\u002Fdev\u002Fvideo15 \\\n  --device \u002Fdev\u002Fvideo16:\u002Fdev\u002Fvideo16 \\\n  --device \u002Fdev\u002Fvideo17:\u002Fdev\u002Fvideo17 \\\n  --device \u002Fdev\u002Fvideo18:\u002Fdev\u002Fvideo18 \\\n  --device \u002Fdev\u002Fvideo19:\u002Fdev\u002Fvideo19 \\\n  --device \u002Fdev\u002Fvideo20:\u002Fdev\u002Fvideo20 \\\n  --device \u002Fdev\u002Fvideo21:\u002Fdev\u002Fvideo21 \\\n  --device \u002Fdev\u002Fvideo22:\u002Fdev\u002Fvideo22 \\\n  --device \u002Fdev\u002Fvideo23:\u002Fdev\u002Fvideo23 \\\n  --device \u002Fdev\u002Fvideo24:\u002Fdev\u002Fvideo24 \\\n  --device \u002Fdev\u002Fvideo25:\u002Fdev\u002Fvideo25 \\\n  --device \u002Fdev\u002Fvideo26:\u002Fdev\u002Fvideo26 \\\n  --device \u002Fdev\u002Fvideo27:\u002Fdev\u002Fvideo27 \\\n  --device \u002Fdev\u002Fvideo28:\u002Fdev\u002Fvideo28 \\\n  --device \u002Fdev\u002Fvideo29:\u002Fdev\u002Fvideo29 \\\n  --device \u002Fdev\u002Fvideo30:\u002Fdev\u002Fvideo30 \\\n  --device \u002Fdev\u002Fvideo31:\u002Fdev\u002Fvideo31 \\\n  --device \u002Fdev\u002Fvideo32:\u002Fdev\u002Fvideo32 \\\n  --device \u002Fdev\u002Fvideo33:\u002Fdev\u002Fvideo33 \\\n  --device \u002Fdev\u002Fvideo34:\u002Fdev\u002Fvideo34 \\\n  --device \u002Fdev\u002Fvideo35:\u002Fdev\u002Fvideo35 \\\n  --device \u002Fdev\u002Fvideo36:\u002Fdev\u002Fvideo36 \\\n  --device \u002Fdev\u002Fvideo37:\u002Fdev\u002Fvideo37 \\\n  --device \u002Fdev\u002Fvideo38:\u002Fdev\u002Fvideo38 \\\n  --device \u002Fdev\u002Fvideo39:\u002Fdev\u002Fvideo39 \\\n  --device \u002Fdev\u002Fvideo40:\u002Fdev\u002Fvideo40 \\\n  --device \u002Fdev\u002Fvideo41:\u002Fdev\u002Fvideo41 \\\n  --device \u002Fdev\u002Fvideo42:\u002Fdev\u002Fvideo42 \\\n  --device \u002Fdev\u002Fvideo43:\u002Fdev\u002Fvideo43 \\\n  --device \u002Fdev\u002Fvideo44:\u002Fdev\u002Fvideo44 \\\n  --device \u002Fdev\u002Fvideo45:\u002Fdev\u002Fvideo45 \\\n  --device \u002Fdev\u002Fvideo46:\u002Fdev\u002Fvideo46 \\\n  --device \u002Fdev\u002Fvideo47:\u002Fdev\u002Fvideo47 \\\n  --device \u002Fdev\u002Fvideo48:\u002Fdev\u002Fvideo48 \\\n  --device \u002Fdev\u002Fvideo49:\u002Fdev\u002Fvideo49 \\\n  --device \u002Fdev\u002Fvideo50:\u002Fdev\u002Fvideo50 \\\n  --device \u002Fdev\u002Fvideo51:\u002Fdev\u002Fvideo51 \\\n  --device \u002Fdev\u002Fvideo52:\u002Fdev\u002Fvideo52 \\\n  --device \u002Fdev\u002Fvideo53:\u002Fdev\u002Fvideo53 \\\n  --device \u002Fdev\u002Fvideo54:\u002Fdev\u002Fvideo54 \\\n  --device \u002Fdev\u002Fvideo55:\u002Fdev\u002Fvideo55 \\\n  --device \u002Fdev\u002Fvideo56:\u002Fdev\u002Fvideo56 \\\n  --device \u002Fdev\u002Fvideo57:\u002Fdev\u002Fvideo57 \\\n  --device \u002Fdev\u002Fvideo58:\u002Fdev\u002Fvideo58 \\\n  --device \u002Fdev\u002Fvideo59:\u002Fdev\u002Fvideo59 \\\n  --device \u002Fdev\u002Fvideo60:\u002Fdev\u002Fvideo60 \\\n  --device \u002Fdev\u002Fvideo61:\u002Fdev\u002Fvideo61 \\\n  --device \u002Fdev\u002Fvideo62:\u002Fdev\u002Fvideo62 \\\n  --device \u002Fdev\u002Fvideo63:\u002Fdev\u002Fvideo63 \\\n  dustynv\u002Fjetson-inference:r36.2.0\n```\n\n*简化版（仅映射必要设备和目录，视具体摄像头数量调整）：*\n```bash\ndocker run --runtime nvidia -it --rm --net=host \\\n  -e DISPLAY=$DISPLAY \\\n  -v \u002Ftmp\u002F.X11-unix:\u002Ftmp\u002F.X11-unix \\\n  -v $(pwd):\u002Fworkspace \\\n  --device \u002Fdev\u002Fvideo0 \\\n  dustynv\u002Fjetson-inference:r36.2.0\n```\n\n> 若需从源码构建，请参考官方文档 `docs\u002Fbuilding-repo-2.md`。\n\n## 3. 基本使用\n\n进入容器后，即可使用内置的命令行工具或编写 Python\u002FC++ 代码。以下是几个核心功能的快速示例。\n\n### 3.1 图像分类 (Image Classification)\n使用预训练的 ResNet-18 模型对单张图片进行分类：\n\n```bash\nimagenet-console images\u002Fgrape.jpg output\u002Fgrape.jpg\n```\n\n实时摄像头分类演示：\n```bash\nimagenet-camera\n```\n\n### 3.2 物体检测 (Object Detection)\n检测图片中的物体（如人、车等）：\n\n```bash\ndetectnet-console images\u002Fpedestrian.jpg output\u002Fpedestrian.jpg\n```\n\n实时摄像头检测演示：\n```bash\ndetectnet-camera\n```\n\n### 3.3 语义分割 (Semantic Segmentation)\n对图像进行像素级分类：\n\n```bash\nsegnet-console images\u002Fsunflower.jpg output\u002Fsunflower.jpg\n```\n\n### 3.4 Python 开发示例\n在容器内创建 `test_inference.py`，使用 Python API 进行图像分类：\n\n```python\nimport jetson.inference\nimport jetson.utils\n\n# 加载预训练模型\nnet = jetson.inference.imageNet(\"resnet18\")\n\n# 加载图像\nimg = jetson.utils.loadImage(\"images\u002Fgrape.jpg\")\n\n# 执行推理\nclass_id, confidence, class_name = net.Classify(img)\n\nprint(f\"识别结果：{class_name} (置信度：{confidence:.2f})\")\n```\n\n运行脚本：\n```bash\npython3 test_inference.py\n```\n\n### 3.5 迁移学习 (训练自定义模型)\n项目支持在 Jetson 上直接使用 PyTorch 收集数据并训练模型。例如，收集猫狗数据集并重新训练分类器：\n\n```bash\n# 启动数据采集工具 (需连接摄像头)\ncollect-classification-dataset.py --dataset my-cat-dog-dataset\n\n# 训练模型\ntrain-classification.py --dataset my-cat-dog-dataset --model my-model.pth\n```\n\n训练完成后，可直接将生成的模型转换为 TensorRT 引擎进行高速推理。\n\n---\n更多高级功能（如姿态估计 `poseNet`、动作识别 `actionNet`、WebRTC 推流等）请参考项目官方文档中的 [API Reference](https:\u002F\u002Frawgit.com\u002Fdusty-nv\u002Fjetson-inference\u002Fmaster\u002Fdocs\u002Fhtml\u002Findex.html) 及 [Code Examples](docs\u002Faux-streaming.md)。","某智慧农业团队需要在温室大棚的嵌入式设备上部署实时病虫害监测系统，以自动识别作物叶片上的病斑并触发警报。\n\n### 没有 jetson-inference 时\n- **部署门槛极高**：开发者需手动配置复杂的 TensorRT 环境，编写大量 C++\u002FCUDA 代码才能将 PyTorch 模型转换为边缘设备可运行的格式，耗时数周。\n- **实时性难以保证**：未经优化的深度学习模型在 Jetson 设备上推理延迟高，无法处理高清摄像头传入的连续视频流，导致漏检严重。\n- **功能开发重复造轮子**：每实现一个新的视觉任务（如从分类切换到目标检测），都需要重新编写底层数据预处理和后处理逻辑，缺乏统一接口。\n- **硬件资源浪费**：由于缺乏针对 NVIDIA GPU 的深度优化，设备算力利用率低，同时占用过多 CPU 资源影响系统其他进程。\n\n### 使用 jetson-inference 后\n- **一键部署加速**：利用内置的 `imageNet` 和 `detectNet`  primitives，团队通过几行 Python 代码即可调用经 TensorRT 优化的预训练模型，将部署周期缩短至几天。\n- **流畅实时推理**：工具自动管理 GPU 内存与流水线，在 Jetson Orin 上实现了每秒 30+ 帧的病虫害检测，确保监控无死角。\n- **模块化快速迭代**：借助统一的 API 接口，开发人员轻松在同一套代码框架下切换图像分类、语义分割等功能，无需重写底层逻辑。\n- **极致性能释放**：基于 TensorRT 的后端自动融合算子并量化模型，显著降低延迟与功耗，让低功耗嵌入式设备也能跑通复杂算法。\n\njetson-inference 通过将复杂的深度学习部署流程标准化与自动化，让开发者能专注于业务逻辑而非底层优化，真正实现了\"Hello AI World\"般的便捷落地。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fdusty-nv_jetson-inference_a68fe911.jpg","dusty-nv","Dustin Franklin","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Fdusty-nv_9e77b1e6.jpg","@NVIDIA Jetson Developer","NVIDIA",null,"dustinf@nvidia.com","https:\u002F\u002Fdeveloper.nvidia.com\u002Fjetson","https:\u002F\u002Fgithub.com\u002Fdusty-nv",[83,87,91,95,99,103,107,110,114,118],{"name":84,"color":85,"percentage":86},"C++","#f34b7d",55.6,{"name":88,"color":89,"percentage":90},"Python","#3572A5",14.5,{"name":92,"color":93,"percentage":94},"CSS","#663399",13.5,{"name":96,"color":97,"percentage":98},"Shell","#89e051",5.8,{"name":100,"color":101,"percentage":102},"Cuda","#3A4E3A",3.9,{"name":104,"color":105,"percentage":106},"HTML","#e34c26",2.3,{"name":108,"color":109,"percentage":32},"JavaScript","#f1e05a",{"name":111,"color":112,"percentage":113},"C","#555555",1.1,{"name":115,"color":116,"percentage":117},"CMake","#DA3434",1,{"name":119,"color":120,"percentage":121},"Dockerfile","#384d54",0.3,8800,3094,"2026-04-13T11:51:51","MIT",4,"Linux","必需 NVIDIA GPU (NVIDIA Jetson 系列设备，如 Nano, Xavier, Orin)，需支持 CUDA 和 TensorRT","未说明 (取决于具体 Jetson 开发板型号)",{"notes":131,"python":132,"dependencies":133},"该工具专为 NVIDIA Jetson 嵌入式设备设计。系统需安装 NVIDIA JetPack SDK (最新支持 JetPack 6)。支持通过 Docker 容器运行或从源码编译。主要功能包括图像分类、物体检测、语义分割、姿态估计等，并利用 TensorRT 进行推理加速，利用 PyTorch 进行模型训练。","未说明 (支持 Python 和 C++)",[134,135,136,137],"TensorRT","PyTorch","CUDA","JetPack",[14,139,15,140],"其他","视频",[142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161],"deep-learning","inference","computer-vision","embedded","image-recognition","object-detection","segmentation","jetson","jetson-tx1","jetson-tx2","jetson-xavier","nvidia","tensorrt","digits","caffe","video-analytics","robotics","machine-learning","jetson-nano","jetson-xavier-nx","2026-03-27T02:49:30.150509","2026-04-14T12:31:39.111639",[165,170,175,180,185,190],{"id":166,"question_zh":167,"answer_zh":168,"source_url":169},32833,"使用自定义数据集训练 SSD-Mobilenet 模型时出现 \"TypeError: unsupported format string passed to PosixPath.__format__\" 错误怎么办？","该错误通常发生在 Python 3.6 环境下，由于代码尝试直接格式化 PosixPath 对象导致。虽然根本修复需要更新库代码，但用户可以检查数据集路径结构是否正确。确保数据目录包含标准的 VOC 格式子文件夹：'Annotations'、'ImageSets'（其中包含 Main 子文件夹及 train.txt 等列表文件）和 'JPEGImages'。此外，如果遇到 'UnicodeDecodeError: ascii codec can't decode byte' 错误，可能是因为标签文件或图片路径中包含非 ASCII 字符，需确保所有文本文件编码正确或移除特殊字符。","https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fissues\u002F1370",{"id":171,"question_zh":172,"answer_zh":173,"source_url":174},32834,"如何在 Jetson TX1\u002FTX2 上运行自己训练的 Caffe 模型（如通过 NVIDIA DIGITS 训练的模型）？","首先将训练好的模型文件（deploy.prototxt, .caffemodel, mean.binaryproto, labels.txt）复制到 jetson-inference 的 data\u002Fnetworks\u002F 目录下。然后修改源码（如 imagenet-console.cpp），使用 imageNet::Create 函数加载这些文件路径，例如：imageNet* net = imageNet::Create(\"networks\u002Fsamplemodel\u002Fdeploy.prototxt\", \"networks\u002Fsamplemodel\u002Fsnapshot_iter_2430.caffemodel\", \"networks\u002Fsamplemodel\u002Fmean.binaryproto\", \"networks\u002Fsamplemodel\u002Flabels.txt\");。修改后重新编译项目。注意输出层名称必须匹配，如果报错 \"failed to retrieve tensor for output 'prob'\"，请检查 deploy.prototxt 中的输出层名称是否为 'prob' 或相应分类层的名称。","https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fissues\u002F71",{"id":176,"question_zh":177,"answer_zh":178,"source_url":179},32835,"在 Jetson Nano 上训练模型时提示 \"Module torch not found\" 且安装脚本卡住怎么办？","这通常是因为构建 Jetson-inference 时跳过了 PyTorch 的安装。您可以尝试运行项目根目录下的 .\u002Finstall-pytorch.sh 脚本进行单独安装。如果在运行脚本时看到选项列表但安装过程日志无限滚动或卡住，可能是网络问题或依赖冲突。建议先确认 JetPack 版本（如 4.3）与 PyTorch 版本的兼容性。如果自动脚本失败，可以尝试手动安装适配 Jetson 的 PyTorch whl 包，或者在构建项目时确保网络连接稳定以完成依赖下载。","https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fissues\u002F1563",{"id":181,"question_zh":182,"answer_zh":183,"source_url":184},32836,"同时连接两个 USB 摄像头运行物体检测导致 Jetson 过热崩溃如何解决？","同时运行两个摄像头会显著增加 CPU\u002FGPU 负载和功耗，导致散热不足而崩溃。解决方案包括：1. 加强散热，务必加装风扇主动散热，仅靠被动散热片可能不足以应对双路视频流；2. 升级电源，使用更高电流（如 5V 4A）的电源适配器，避免供电不足；3. 优化代码，不要简单复制代码块运行两个独立循环，应合并处理逻辑或使用多线程\u002F多进程平衡负载；4. 降低分辨率或帧率，减少单个摄像头的输入数据量以减轻处理压力。","https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fissues\u002F882",{"id":186,"question_zh":187,"answer_zh":188,"source_url":189},32837,"在使用 TensorRT 3.0 构建 SSD 模型时遇到 \"Plugin layer output count is not equal to caffe output count\" 错误如何处理？","这是因为 TensorRT 3.0 官方插件库中可能缺少或不完全支持 SSD 所需的 DetectionOutput 层实现。解决方法是手动实现 DetectionOutput 层插件。用户需要编写自定义的 Plugin 代码来替代库中缺失的功能，确保输出维度与 Caffe 模型定义的输出计数一致。如果无法自行开发，可以暂时移除 prototxt 文件中的 DetectionOutput 层进行测试，但这会导致无法直接获得最终的检测框结果，仅能获取中间层（如 mbox_loc, mbox_conf_flatten）的输出。","https:\u002F\u002Fgithub.com\u002Fdusty-nv\u002Fjetson-inference\u002Fissues\u002F171",{"id":191,"question_zh":192,"answer_zh":193,"source_url":184},32838,"程序启动前日志级别解析导致的调试信息过多或报错如何处理？","如果在程序解析命令行参数（如 --log-level）之前就输出了日志，可能导致日志级别设置不生效。维护者已将相关日志调用降级为 LogDebug() 以避免干扰。如果您遇到类似问题，可以检查代码中是否在初始化配置前调用了日志函数。对于用户而言，确保使用最新的代码版本，其中包含了针对此问题的修复提交（commit 5718df1495859808d5d076efea597b7e641bcb96），这样可以在解析参数后正确控制日志输出级别。",[195],{"id":196,"version":197,"summary_zh":198,"released_at":199},247555,"model-mirror-190618","本发布包含该仓库所使用的 DNN 模型的镜像下载。\n\n在构建过程中，`jetson-inference` 仓库会自动尝试为您下载这些模型。\n\n模型的主要存储位置位于 Box.com 上。然而，来自中国的用户可能无法访问 Box.com，因此，如果您的系统无法从其主位置下载模型，您可以从下方获取所需模型。\n\n对于您希望在运行时用于推理的每个模型，请将下方对应的归档文件下载到您的 `\u003Cjetson-inference>\u002Fdata\u002Fnetworks` 目录中，然后使用以下命令解压：\n\n```bash\ncd \u003Cjetson-inference>\u002Fdata\u002Fnetworks\u002F\ntar -zxvf \u003Cmodel-archive-name>.tar.gz\n```\n\n在此之前，请按常规方式克隆该仓库。本发布页面并非用于存放代码版本。","2019-06-18T18:27:49"]