feat(Pose): Add YOLO Pose Estimation inference module with C++ and Python bindings, include example (#58)

close #58

Author: laugh12321

examples/pose/CMakeLists.txt

@@ -0,0 +1,86 @@
+# 设置项目
+cmake_minimum_required(VERSION 3.15.0)
+cmake_policy(SET CMP0091 NEW)
+cmake_policy(SET CMP0146 OLD)
+project(pose LANGUAGES CXX CUDA)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+# 设置 C++ 标准
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# 添加依赖项
+find_package(CUDA REQUIRED)
+set(CMAKE_CUDA_ARCHITECTURES native)
+set(CUDA_PATH ${CUDA_TOOLKIT_ROOT_DIR})
+
+find_package(OpenCV REQUIRED)
+
+# 添加编译选项
+option(TENSORRT_PATH "TensorRT Path. Example: /usr/local/tensorrt" "")
+if(NOT DEFINED TENSORRT_PATH)
+    message(FATAL_ERROR "TensorRT path is not set. Please specify the TensorRT path.")
+endif()
+
+option(DEPLOY_PATH "TensorRT-YOLO Project Path." "")
+if(NOT DEFINED DEPLOY_PATH)
+    message(FATAL_ERROR "TensorRT-YOLO project path is not set. Please specify the TensorRT-YOLO Project path.")
+endif()
+
+# 添加可执行文件
+add_executable(pose "")
+
+# 包含头文件目录
+target_include_directories(pose PRIVATE
+    ${CUDA_INCLUDE_DIRS}
+    ${TENSORRT_PATH}/include
+    ${OpenCV_INCLUDE_DIRS}
+    ${DEPLOY_PATH}/include
+)
+
+# 链接库目录
+target_link_directories(pose PRIVATE
+    ${TENSORRT_PATH}/lib
+    ${DEPLOY_PATH}/lib
+)
+
+# 私有源文件
+target_sources(pose PRIVATE pose.cpp)
+
+# 私有编译定义
+target_compile_definitions(pose PRIVATE ${CUDA_DEFINITIONS})
+
+# 私有链接库
+target_link_libraries(pose PRIVATE
+    ${CUDA_cudart_LIBRARY}
+    ${OpenCV_LIBS}
+    deploy
+)
+
+# 根据平台选择不同的 TensorRT 库
+if(MSVC AND EXISTS ${TENSORRT_PATH}/lib/nvinfer_10.dll)
+    target_link_libraries(pose PRIVATE nvinfer_10 nvinfer_plugin_10 nvonnxparser_10)
+else()
+    target_link_libraries(pose PRIVATE nvinfer nvinfer_plugin nvonnxparser)
+endif()
+
+# 设置输出目录
+set_target_properties(pose PROPERTIES
+    OUTPUT_NAME "pose"
+    RUNTIME_OUTPUT_DIRECTORY "${CMAKE_SOURCE_DIR}/bin"
+)
+
+# 编译选项
+if(MSVC)
+    target_compile_options(pose PRIVATE
+        $<$<CONFIG:Release>:-O2>
+    )
+    set_property(TARGET pose PROPERTY MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>")
+else()
+    target_compile_options(pose PRIVATE
+        $<$<COMPILE_LANGUAGE:CXX>:-O3 -flto=auto>
+    )
+    target_link_options(pose PRIVATE
+        $<$<COMPILE_LANGUAGE:CXX>:-O3 -flto=auto>
+    )
+endif()

examples/pose/README.en.md

@@ -0,0 +1,97 @@
+[简体中文](README.md) | English
+
+# Pose Estimation Inference Example
+
+This example uses the yolo11n-pose model to demonstrate how to perform pose estimation inference using the Command Line Interface (CLI), Python, and C++.
+
+[yolo11n-pose.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/yolo11n-pose.pt)，[【TestImages】COCO-Pose-part.zip](https://www.ilanzou.com/s/kBby4w1D)
+
+Please download the required `yolo11n-pose.pt` model file and test images through the provided link, and save the model file to the `models` folder, and place the extracted test images into the `images` folder after unzipping.
+
+## Model Export
+
+> [!IMPORTANT]
+>
+> If you only want to export the ONNX model (with TensorRT plugins) that can be used for inference in this project through the `tensorrt_yolo` provided Command Line Interface (CLI) tool `trtyolo`, you can install it via [PyPI](https://pypi.org/project/tensorrt-yolo) by simply executing the following command:
+>
+> ```bash
+> pip install -U tensorrt_yolo
+> ```
+> 
+> If you want to experience the same inference speed as C++, please refer to [Install-tensorrt_yolo](../../docs/en/build_and_install.md#install-tensorrt_yolo) to build the latest version of `tensorrt_yolo` yourself.
+
+Use the following command to export the ONNX format with the [EfficientRotatedNMS](../../plugin/efficientRotatedNMSPlugin/) plugin. For detailed `trtyolo` CLI export methods, please read [Model Export](../../docs/en/model_export.md):
+
+```bash
+trtyolo export -w models/yolo11n-pose.pt -v yolo11 -o models -s
+```
+
+After running the above command, a `yolo11n-pose.onnx` file with a `batch_size` of 1 will be generated in the `models` folder. Next, use the `trtexec` tool to convert the ONNX file to a TensorRT engine (fp16):
+
+```bash
+trtexec --onnx=models/yolo11n-pose.onnx --saveEngine=models/yolo11n-pose.engine --fp16 --staticPlugins=/path/to/your/TensorRT-YOLO/lib/plugin/libcustom_plugins.so --setPluginsToSerialize=/path/to/your/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
+```
+
+## Model Inference
+
+> [!IMPORTANT]
+>
+> The `tensorrt_yolo` installed via [PyPI](https://pypi.org/project/tensorrt-yolo) only provides the ONNX model (with TensorRT plugins) for inference in this project and does not provide inference capabilities.
+> If you want to experience the same inference speed as C++, please refer to [Install-tensorrt_yolo](../../docs/en/build_and_install.md#install-tensorrt_yolo) to build the latest version of `tensorrt_yolo` yourself.
+
+### Inference Using CLI
+
+> [!NOTE] 
+> The `--cudaGraph` command added from version 4.0 can further accelerate the inference process, but this feature only supports static models.
+> 
+> From version 4.3 and later, support for pose estimation inference is added. The command `-m 2, --mode 2` is used to select the pose estimation.
+
+1. Use the `trtyolo` command-line tool from the `tensorrt_yolo` library for inference. Run the following command to view help information:
+
+    ```bash
+    trtyolo infer --help
+    ```
+
+2. Run the following command for inference:
+
+    ```bash
+    trtyolo infer -e models/yolo11n-pose.engine -m 3 -i images -o output -l labels.txt --cudaGraph
+    ```
+
+    The inference results will be saved in the `output` folder, and a visualization result will be generated.
+
+### Inference Using Python
+
+1. Use the `tensorrt_yolo` library to run the example script `pose.py` for inference.
+2. Run the following command for inference:
+
+    ```bash
+    python pose.py -e models/yolo11n-pose.engine -i images -o output -l labels.txt --cudaGraph
+    ```
+
+### Inference Using C++
+
+1. Ensure that the project has been compiled according to the [`TensorRT-YOLO` Compilation](../../docs/en/build_and_install.md#tensorrt-yolo-compile).
+2. Compile `pose.cpp` into an executable:
+
+    ```bash
+    # Compile using xmake
+    xmake f -P . --tensorrt="/path/to/your/TensorRT" --deploy="/path/to/your/TensorRT-YOLO"
+    xmake -P . -r
+
+    # Compile using cmake
+    mkdir -p build && cd build
+    cmake -DTENSORRT_PATH="/path/to/your/TensorRT" -DDEPLOY_PATH="/path/to/your/TensorRT-YOLO" .. 
+    cmake --build . -j8 --config Release
+    ```
+
+    After compilation, the executable file will be generated in the `bin` folder of the project root directory.
+
+3. Run the following command for inference:
+
+    ```bash
+    cd bin
+    ./pose -e ../models/yolo11n-pose.engine -i ../images -o ../output -l ../labels.txt --cudaGraph
+    ```
+
+Through the above methods, you can successfully complete model inference.

examples/pose/README.md

@@ -0,0 +1,97 @@
+[English](README.en.md) | 简体中文
+
+# 姿态识别推理示例
+
+本示例以 yolo11n-pose 模型为例，展示如何使用命令行界面（CLI）、Python 和 C++ 三种方式进行姿态识别推理。
+
+[yolo11n-pose.pt](https://github.com/ultralytics/assets/releases/download/v8.3.0/yolo11n-pose.pt)，[【测试图片】COCO-Pose-part.zip](https://www.ilanzou.com/s/kBby4w1D)
+
+请通过提供的链接下载所需的 `yolo11n-pose.pt` 模型文件和测试图片，并将模型文件保存至 `models` 文件夹，测试图片解压后存放至 `images` 文件夹。
+
+## 模型导出
+
+> [!IMPORTANT]
+>
+> 如果您仅想通过 `tensorrt_yolo` 提供的命令行界面（CLI）工具 `trtyolo`，导出可供该项目推理的 ONNX 模型（带 TensorRT 插件），可以通过 [PyPI](https://pypi.org/project/tensorrt-yolo) 安装，只需执行以下命令即可：
+>
+> ```bash
+> pip install -U tensorrt_yolo
+> ```
+> 
+> 如果想体验与 C++ 同样的推理速度，则请参考 [安装-tensorrt_yolo](../../docs/cn/build_and_install.md#安装-tensorrt_yolo) 自行构建最新版本的 `tensorrt_yolo`。
+
+使用以下命令导出带 [EfficientRotatedNMS](../../plugin/efficientRotatedNMSPlugin/)  插件的 ONNX 格式，详细的 `trtyolo` CLI 导出方法请阅读 [模型导出](../../docs/cn/model_export.md)：
+
+```bash
+trtyolo export -w models/yolo11n-pose.pt -v yolo11 -o models -s
+```
+
+运行上述命令后，`models` 文件夹中将生成一个 `batch_size` 为 1 的 `yolo11n-pose.onnx` 文件。接下来，使用 `trtexec` 工具将 ONNX 文件转换为 TensorRT 引擎（fp16）：
+
+```bash
+trtexec --onnx=models/yolo11n-pose.onnx --saveEngine=models/yolo11n-pose.engine --fp16 --staticPlugins=/path/to/your/TensorRT-YOLO/lib/plugin/libcustom_plugins.so --setPluginsToSerialize=/path/to/your/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
+```
+
+## 模型推理
+
+> [!IMPORTANT]
+>
+> 通过 [PyPI](https://pypi.org/project/tensorrt-yolo) 安装的 `tensorrt_yolo` 仅提供可供该项目推理的 ONNX 模型（带 TensorRT 插件） 功能，不提供推理功能。
+> 如果想体验与 C++ 同样的推理速度，则请参考 [安装-tensorrt_yolo](../../docs/cn/build_and_install.md#安装-tensorrt_yolo) 自行构建最新版本的 `tensorrt_yolo`。
+
+### 使用 CLI 进行推理
+
+> [!NOTE] 
+> 从 4.0 版本开始新增的 `--cudaGraph` 指令可以进一步加速推理过程，但该功能仅支持静态模型。
+> 
+> 从 4.3 以后的版本开始，支持姿态识别，指令 `-m 3, --mode 3` 用于选择姿态识别。
+
+1. 使用 `tensorrt_yolo` 库的 `trtyolo` 命令行工具进行推理。运行以下命令查看帮助信息：
+
+    ```bash
+    trtyolo infer --help
+    ```
+
+2. 运行以下命令进行推理：
+
+    ```bash
+    trtyolo infer -e models/yolo11n-pose.engine -m 3 -i images -o output -l labels.txt --cudaGraph
+    ```
+
+    推理结果将保存至 `output` 文件夹，并生成可视化结果。
+
+### 使用 Python 进行推理
+
+1. 使用 `tensorrt_yolo` 库运行示例脚本 `pose.py`进行推理。
+2. 运行以下命令进行推理：
+
+    ```bash
+    python pose.py -e models/yolo11n-pose.engine -i images -o output -l labels.txt --cudaGraph
+    ```
+
+### 使用 C++ 进行推理
+
+1. 确保已按照 [`TensorRT-YOLO` 编译](../../docs/cn/build_and_install.md##rensorrt-yolo-编译) 对项目进行编译。
+2. 将 `pose.cpp` 编译为可执行文件：
+
+    ```bash
+    # 使用 xmake 编译
+    xmake f -P . --tensorrt="/path/to/your/TensorRT" --deploy="/path/to/your/TensorRT-YOLO"
+    xmake -P . -r
+
+    # 使用 cmake 编译
+    mkdir -p build && cd build
+    cmake -DTENSORRT_PATH="/path/to/your/TensorRT" -DDEPLOY_PATH="/path/to/your/TensorRT-YOLO" .. 
+    cmake --build . -j8 --config Release
+    ```
+
+    编译完成后，可执行文件将生成在项目根目录的 `bin` 文件夹中。
+
+3. 使用以下命令运行推理：
+
+    ```bash
+    cd bin
+    ./pose -e ../models/yolo11n-pose.engine -i ../images -o ../output -l ../labels.txt --cudaGraph
+    ```
+
+通过以上方式，您可以顺利完成模型推理。

examples/pose/images/.keep

@@ -0,0 +1 @@
+person