How barrydoooit/mwpose3d works
Project type and purpose
• This is a research-oriented full-stack Python package (“mwpose3d”) for 3D human-pose estimation on mmWave radar point clouds.
• It provides:
• A core library (mwpose3d/) containing datasets, models, runners, hooks, utils, and visualization.
• A configurable training & evaluation pipeline built on OpenMMLab’s mmengine.
• Application scripts (tools/) and GUI apps (apps/) for dataset collection and online estimation.
• Project templates (projects/) with example configs and models.
File structure
• Top-level directories:
• mwpose3d/: core library implementing data pipelines, model definitions, training/inference loops, evaluation, visualization, and a registry for plugins.
• configs/: default and app-specific configuration files.
• apps/: GUI applications for dataset collection and online/offline pose estimation.
• tools/: command-line scripts for data conversion, dataset inspection, training, evaluation, and containerization.
• projects/: sample research projects, each with its configs, models, and entry point.
• Patterns & frameworks:
• Plugin registry (mwpose3d/registry.py).
• Config-driven instantiation via mmengine.
• Hook-based training/inference loops (mwpose3d/runner/hooks).
• Modular dataset transforms (mwpose3d/datasets/transforms).
• Clear separation of core library and application layers.
Key README insights
• Emphasizes modularity, extensibility, and use of mmengine’s config system.
• Describes workflows: dataset preparation, training & evaluation, online estimation, and custom GUI apps.
Diagramming recommendations
• Blue: Frontends (CLI/GUI)
• Green: Core modules (data, model, runner)
• Orange: Extensions (CUDA, external services)
• Gray: Configuration & registry