mv-api

mv-api

Raw HoCap full exo/ego pipeline port for the examples monorepo.

The package scope is CLI/API only. Gradio and client UIs are intentionally out of scope for the first port.

This package is intended to be run from the Pixi workspace. Its model packages (monopriors, wilor-nano, and rtmlib) are resolved by the root pixi.toml path/git sources rather than standalone PyPI releases.

Run

The full pipeline is exposed through the node app entrypoint:

pixi run -e mv-api mv-api-full-app --rr-config.save artifacts/node-validation/hocap-node-app.rrd --rr-config.headless --max-frames 1 --camera-source dataset

Catalog Notes

The catalog tools (mv-api-catalog-prediction-layer and the macOS mv-api-catalog-register) connect to an already-running Rerun catalog via CatalogClient — they no longer mount the catalog in-process. Before running either, start the base catalog with the simplecv tasks:

pixi run -e simplecv-catalog --frozen simplecv-catalog-serve       # tier 1: serve the catalog (rerun server)
pixi run -e simplecv-catalog --frozen simplecv-catalog-register    # tier 2: register the base ExoEgo Forge recordings

Both tools default to --catalog-url rerun+http://127.0.0.1:9988. Prediction layers are attached with on_duplicate=REPLACE, so re-running is idempotent.

The prediction dataloader samples each exo stream at its native frame rate via the stock RerunIterableDataset + FixedRateSampling — sub-native decimation is avoided because the sampling grid drops reference packets and can make AV1 decode fail or return wrong pixels (RR-5087). See docs/catalog_dataloader.md.