A unified, plugin-oriented framework for training and evaluating pose estimation models. I have built this to solve the reproducibility problem in the field and to perform testing for my publications during my PhD.
Architecture diagram will be here (at some point)
Motivation
Researchers in human pose estimation typically release their own codebases, each with different data formats, training loops, and evaluation protocols. Comparing models fairly — or retraining them on a new dataset — requires substantial engineering effort each time. This framework was built to eliminate that overhead.
Highlights
Plugin architecture
Models, datasets, and augmentation strategies are registered as plugins via YAML config. Swapping a model or dataset requires changing a single line. Every new addition (model, dataset, transform etc.) can be added in a separate folder and registered so no refactoring is needed.
Unified model zoo
2D: HRNet, OpenPose-Torch, SimpleBaseline2D. 3D: SimpleBaseline3D, VideoPose3D, MotionAGFormer, KTPFormer, Context-Aware Pose Former, GASTNet.
Automated evaluation pipeline
All experiment outputs (metrics, 2d and 3D poses, checkpoints) are exported to HDF5. A report generator produces a self-contained HTML file with plots, tables, and per-sequence breakdowns.
Reproducibility-first
Every run is fully described by its YAML config. Configs are versioned alongside checkpoints, making experiments reproducible.
Impact
Used internally at the lab in multiple research projects and in an industry collaboration for clinical markerless motion analysis.