openpose: A Simple Pose Estimation CLI Tool

openpose is a command-line interface (CLI) tool for pose estimation, built with OpenCV. It processes input images to detect poses including body, hand, face, or all.

Installation

Prerequisites

• OpenCV (compiled with C++17 standard support)
• CMake
• A C++ compiler supporting C++17 (e.g., g++ or clang++)

Build Steps

1. Navigate to the project root directory.
2. Create and enter the build directory:

   mkdir build && cd build
   

3. Configure the build:

   cmake ..
   

4. Build the project:

   make
   

The executable will be available at ./build/openpose.

Usage

Run the tool with the following command structure:

Usage: ./build/openpose --input <file> --output <file> [OPTIONS]
Options:
  -i, --input <file>         Required. Path to the input image.
  -o, --output <file>        Optional. Path to save the output image (default: unified_output.jpg).
  -t, --threshold <value>    Pose keypoint detection threshold (0.0-1.0, default: 0.1).
  --person-threshold <value> Person detection confidence threshold for multi-person (default: 0.5).
  --nms-threshold <value>    Non-Maximum Suppression threshold for multi-person (default: 0.4).
  -b, --blend <factor>       Original image opacity (0.0-1.0, default: 0.5).
  -m, --models <directory>   Models base directory (default: ./models).
  -M, --mode <modes>         Modes: comma-separated list of 'body', 'hand', 'face', 'all' (default: body).
  -v, --verbose              Enable verbose output.
  --single-person            Disable multi-person detection (multi-person is on by default).
  --no-validate-connectivity Deactivates filtering of noisy points (on by default).
  --no-interpolate           Deactivates estimation of missing joints (on by default).
  --draw-foot                Draw the foot keypoints (off by default).
  --no-rainbow               Disable rainbow coloring for skeletons.
  -h, --help                 Display this help message.

Required Arguments

• -i, --input <file>: Path to the input image to process.

Optional Arguments

• -o, --output <file>: Path to save the output image (default: unified_output.jpg).
• -t, --threshold <value>: Pose keypoint detection threshold (0.0-1.0, default: 0.1). Higher values make detection more strict.
• --person-threshold <value>: Person detection confidence threshold for multi-person scenarios (default: 0.5).
• --nms-threshold <value>: Non-Maximum Suppression threshold to reduce overlapping detections in multi-person (default: 0.4).
• -b, --blend <factor>: Opacity of the original image in the output (0.0-1.0, default: 0.5). 1.0 = full original, 0.0 = annotations only.
• -m, --models <directory>: Base directory for models (default: ./models).
• -M, --mode <modes>: Comma-separated list of detection modes: body, hand, face, all (default: body).
• -v, --verbose: Enable verbose logging.
• --single-person: Use single-person mode (disables multi-person detection; multi-person is default).
• --no-validate-connectivity: Disable filtering of noisy keypoints (enabled by default).
• --no-interpolate: Disable estimation of missing joints (enabled by default).
• --draw-foot: Enable drawing of foot keypoints (disabled by default).
• --no-rainbow: Disable rainbow colors for skeleton visualization (enabled by default).
• -h, --help: Show the help message and exit.

Dependencies

• OpenCV (for computer vision and image processing)
• C++17 standard (for modern language features)

Models

The tool requires pre-trained models for pose estimation. Download them from: https://huggingface.co/fszontagh/openpose-pose-estimation

Extract the models and place them in the ./models directory (relative to the project root). Ensure the directory structure matches the expected model paths used in the code.