ego-vision 0.1.1

Composable egocentric vision model components for hand and camera geometry pipelines.

ego-vision

Composable egocentric vision model components and pipelines for hand and camera geometry.

The initial scope is intentionally narrow: raw model wrappers plus an episode-level pipeline boundary for the best VGGT-Omega + HaWoR stack.

episodes of av.VideoFrame
  -> VGGTOmegaModel
  -> HaWoRHandReconstructor
       YoloHandDetector + HaWoRModel + ManoLayer
  -> HaWoRHandMotionInfiller
  -> HandTrackingPipeline result

The egovision.models namespace contains raw model components. The egovision.pipelines namespace owns model scheduling and hand-tracking orchestration.

VGGT-Omega and the HaWoR inference modules needed by the pipeline live under egovision.models, so users do not need separate research-repo checkouts or PYTHONPATH setup. Checkpoints are still loaded from local paths or Hugging Face.

from egovision import (
    EpisodeInput,
    HandTrackingConfig,
    HandTrackingPipeline,
    HaworReconstructionConfig,
    VggtOmegaConfig,
)

pipeline = HandTrackingPipeline(
    HandTrackingConfig(
        hand_reconstruction=HaworReconstructionConfig(batch_size=64),
        camera_pose_estimator=VggtOmegaConfig(batch_size=1),
    )
)
results = pipeline.predict_episodes(
    [EpisodeInput(frames=av_frames)]
)

Install

pip install ego-vision

Heavy model dependencies are optional:

pip install "ego-vision[models]"

Development

uv sync --dev
uv run pytest
uv run ruff check --force-exclude .
uv run ruff format --force-exclude --check .
uv run ty check

Local hooks use the same commands:

uv run pre-commit run --all-files

Release

Publishing is handled by GitHub Actions with PyPI trusted publishing.

To release:

1. Bump version in pyproject.toml.
2. Push the commit to main.
3. The Publish package workflow builds the wheel and sdist, publishes to PyPI, then creates the vX.Y.Z tag and GitHub release.

The workflow can also be run manually from GitHub Actions with target testpypi or pypi. The normal release path only needs the pypi trusted publishing environment configured for the ego-vision project.