1 oracle.pipelines.scene_labels

oracle.pipelines.scene_labels

End-to-end oracle RRI label generation pipeline (non-Streamlit).

This module provides a small orchestration layer that is reusable in:

  • the Streamlit dashboard (visualization),
  • offline dataset preprocessing, and
  • training-time online label generation (candidates + oracle RRI).

The goal is to make data-flow explicit and to keep performance controls local to the compute code (chunk sizes, backprojection stride, etc.).

1.1 Classes

Name Description
OracleRriLabeler Compute oracle RRI labels for candidates in a single snippet.
OracleRriLabelerConfig Config-as-factory wrapper for OracleRriLabeler.
OracleRriSample One fully labeled candidate set for an EFM snippet.

1.1.1 OracleRriLabeler

oracle.pipelines.scene_labels.OracleRriLabeler(config)

Compute oracle RRI labels for candidates in a single snippet.

1.1.1.1 Methods

Name Description
run Run the full candidate→render→RRI pipeline for one snippet.
1.1.1.1.1 run
oracle.pipelines.scene_labels.OracleRriLabeler.run(sample)

Run the full candidate→render→RRI pipeline for one snippet.

1.1.1.1.1.1 Parameters
Name Type Description Default
sample EfmSnippetView Input snippet including semi-dense points and GT mesh tensors. required
1.1.1.1.1.2 Returns
Name Type Description
OracleRriSample A batch containing candidates, renders, backprojected point clouds,
OracleRriSample and oracle RRI values.

1.1.2 OracleRriLabelerConfig

oracle.pipelines.scene_labels.OracleRriLabelerConfig()

Config-as-factory wrapper for OracleRriLabeler.

This config composes the existing stage configs (generation, rendering, scoring) and adds a small number of pipeline-level knobs.

1.1.2.1 Attributes

Name Description
target_type Return the runtime pipeline type constructed by this config.
device Geometry tensor device resolved under PYTORCH3D_BACKEND policy.
generator Candidate generation configuration; omitted devices inherit before provider admission.
depth Depth rendering configuration; omitted devices inherit before provider admission.
oracle Oracle RRI scoring configuration.
backprojection_stride Pixel stride used when backprojecting depth maps to point clouds.
verbosity Pipeline progress verbosity; nested stages retain their own settings.

1.1.3 OracleRriSample

oracle.pipelines.scene_labels.OracleRriSample(
    sample,
    candidates,
    depths,
    candidate_pcs,
    rri,
)

One fully labeled candidate set for an EFM snippet.

The payload keeps actor-visible snippet/candidate evidence beside oracle-only mesh renders and RRI labels. Candidate-aligned members share a compact valid-candidate axis C; rows map back to the sampled shell via :meth:CandidateSamplingResult.candidate_shell_indices.

1.1.3.1 Attributes

Name Description
sample Typed snippet view owning observations, calibration, and GT mesh data.
candidates Full-shell provenance plus the compact valid world-from-camera poses.
depths Metric z-depth maps Tensor["C H W", float] rendered from the GT mesh.
candidate_pcs Padded world-frame candidate points Tensor["C P 3", float] in metres.
rri Oracle reconstruction-improvement labels and distance decomposition over C.
Back to top