SynthPose: Clean Architecture Implementation

A robust, modular Python package for human pose estimation using RT-DETR and VitPose. This project has been refactored to follow Clean Architecture principles, making it suitable for research, extension, and integration into larger systems (e.g., Avalonia/C# apps).

Features

• Clean Architecture: Separation of concerns into Domain, Application, Adapters, and Infrastructure layers.
• State-of-the-Art Models:
  • Detection: RT-DETR (PekingU/rtdetr_r50vd_coco_o365)
  • Pose Estimation: VitPose (stanfordmimi/synthpose-vitpose-huge-hf or base)
• Standardized Output: OpenPose-compatible JSON format.
• Configurable: Support for YAML configuration and CLI arguments.
• Type Safe: Fully typed codebase using Python type hints and Pydantic.

Architecture

The project is organized as follows:

• domain/: Core business logic and entities (FrameData, Person, Keypoint). No external dependencies.
• application/: Use cases and orchestration (VideoProcessor). Depends only on Domain.
• adapters/: Implementations of interfaces (OpenCVVideoSource, RTDetrPersonDetector, VitPoseEstimator, OpenPoseResultWriter). Depends on external libraries (torch, cv2).
• infrastructure/: Configuration and Entry points (CLI, Settings).

Installation

This project uses Poetry for dependency management.

# Install dependencies
poetry install

Usage

CLI

Run the processor via the synthpose command (defined in pyproject.toml):

# Basic usage
poetry run synthpose process path/to/video.mp4

# With options
poetry run synthpose process video.mp4 --mode base --device cpu

# Using a config file
poetry run synthpose process video.mp4 --config configs/default.yaml

Configuration

You can use a YAML file to configure the pipeline:

# configs/example.yaml
device: "cuda"
model_mode: "huge"
det_threshold: 0.3
vis_show_weight: true

Output

• Video: {input_name}_output.mp4 (Visualized result)
• JSON: pose/{input_name}_json/{input_name}_{frame_id}.json (Keypoints data)

Testing

poetry run pytest

License

MIT License

References

• VitPose Paper
• SynthPose Model