steps.md

Pipeline steps

The MAMMA local runner orchestrates five pipeline steps. Each step's code is tracked directly under its step directory in this repo; the runner provides a thin builder per step that knows how to invoke the step's entry script with the right argv shape.

Step id	Human name	Step directory	Builder
ma_cap	capture (input)	capture/	inference/steps/ma_cap.py
ma_masks	segmentation	segmentation/	inference/steps/ma_masks.py
ma_2d	landmarks	landmarks/	inference/steps/ma_2d.py
ma_3d	optimization	optimization/	inference/steps/ma_3d.py
ma_vis	visualization	visualization/	inference/steps/ma_vis.py

visualization/ is a library-quality Python module. Importable as from visualization import run_visualization and runnable as python -m visualization. The runner subprocesses visualization/run_ma_vis.py (a thin shim around the same CLI).

Default pipeline

ma_cap ──► ma_masks ──► ma_2d ──► ma_3d ──► ma_vis
                          ▲                    ▲
                          └─── ma_cap ─────────┘

ma_2d consumes both ma_cap (frames) and ma_masks (person masks). ma_3d consumes ma_cap (cameras + frames) and ma_2d (2D landmarks). ma_vis consumes everything upstream and renders the final SMPL-X result overlays.

To run a subset, set enabled: false on the steps you want to skip — the runner drops them and reorders the remaining steps so each one's inputs are produced before it runs.

Adding a step

1. Drop a new builder under inference/steps/<step_id>.py that subclasses StepBuilder and implements python_argv(seq_name).
2. Register it in inference/steps/__init__.py::BUILDERS.
3. Add <step_id> to ALL_STEPS in inference/runner.py.