1 rri_metrics.oracle_rri

rri_metrics.oracle_rri

High-level oracle RRI computation orchestrator.

This module wires together rendering outputs (candidate depth → point clouds), distance primitives (see metrics.py), and configuration defaults to produce per-candidate Relative Reconstruction Improvement (RRI) scores as defined in docs/contents/theory/rri_theory.qmd. The implementation is intentionally kept modular:

Sampling / downsampling is delegated to callers (or future helpers) so that density can be harmonised between P_t and P_{t∪q}.
Distance evaluation is performed via PyTorch3D on GPU for efficiency.
Config-as-factory pattern is used to keep runtime objects serialisable and consistent with the rest of the codebase.

The implemented scalar score is

\[ \mathrm{RRI}(q) = \frac{\Delta(P_t, M) - \Delta(P_t \cup P_q, M)} {\max(\Delta(P_t, M), \epsilon)} \]

where \(\Delta\) is the configured point-mesh error. Target-specific callers pass target-cropped points and meshes; invalid crops raise upstream and must not be silently converted to scene-level labels.

1.1 Classes

Name	Description
OracleRRIConfig	Config-as-factory wrapper for oracle RRI computation.
OracleRRI	Facade to compute oracle RRI for one or more candidates.

# rri_metrics.oracle_rri { #aria_nbv.rri_metrics.oracle_rri }

`rri_metrics.oracle_rri`

High-level oracle RRI computation orchestrator.

This module wires together rendering outputs (candidate depth → point clouds),
distance primitives (see ``metrics.py``), and configuration defaults to produce
per-candidate Relative Reconstruction Improvement (RRI) scores as defined in
``docs/contents/theory/rri_theory.qmd``. The implementation is intentionally
kept modular:

* **Sampling / downsampling** is delegated to callers (or future helpers) so
  that density can be harmonised between ``P_t`` and ``P_{t∪q}``.
* **Distance evaluation** is performed via PyTorch3D on GPU for efficiency.
* **Config-as-factory** pattern is used to keep runtime objects
  serialisable and consistent with the rest of the codebase.

The implemented scalar score is

$$
\mathrm{RRI}(q) =
\frac{\Delta(P_t, M) - \Delta(P_t \cup P_q, M)}
     {\max(\Delta(P_t, M), \epsilon)}
$$

where $\Delta$ is the configured point-mesh error. Target-specific callers pass
target-cropped points and meshes; invalid crops raise upstream and must not be
silently converted to scene-level labels.

## Classes

| Name | Description |
| --- | --- |
| [OracleRRIConfig](aria_nbv.rri_metrics.oracle_rri.OracleRRIConfig.qmd#aria_nbv.rri_metrics.oracle_rri.OracleRRIConfig) | Config-as-factory wrapper for oracle RRI computation. |
| [OracleRRI](aria_nbv.rri_metrics.oracle_rri.OracleRRI.qmd#aria_nbv.rri_metrics.oracle_rri.OracleRRI) | Facade to compute oracle RRI for one or more candidates. |