1 rri_metrics

rri_metrics

Oracle RRI metrics, ordinal labels, and VIN logging utilities.

RRI is computed from point-mesh error before and after adding a candidate view. The directional components are accuracy (point to mesh) and completeness (mesh to point); their scalarized sum is used by the current oracle implementation. For target-aware labels, callers crop points and mesh to the matched target and must mark empty or ambiguous crops invalid rather than assigning low RRI.

CORAL utilities convert continuous RRI-derived supervision into ordered bins for VIN-style one-step scoring. Ordinal predictions are ranking/calibration signals, not geometry metrics by themselves.

1.1 Classes

Name	Description
CoralLayer	CORAL output layer with shared weights and per-threshold biases.
RootEvalPointCloud	Root oracle evaluation point cloud plus lineage metadata.
RriEvaluationPointCloudSource	Source used for the current/root point cloud in oracle RRI labels.
RriRewardMode	Reward used to rank rollout candidates from oracle distance diagnostics.
LabelHistogram	Accumulate label counts for ordinal classes.
Loss	Loss suffixes composed with Stage as {stage}/{loss}.
Metric	Metric suffixes composed with Stage as {stage}/{metric}.
RriErrorStats	Accumulate bias/variance statistics for RRI regression errors.
VinMetrics	Container for VIN metrics computed from candidate rankings.
VinMetricsConfig	Configuration for VIN torchmetrics bundles.
TargetRolloutMetricSummary	Selected-trajectory target-RRI and endpoint metric summary.
RriOrdinalBinner	RRI → ordinal label mapping (CORAL-compatible).
DistanceAggregation	Supported reduction modes for distance tensors.
DistanceBreakdown	Directional distance components used to form Chamfer-style metrics.
RriResult	Batch of per-candidate RRI outcomes and distance diagnostics.

1.2 Functions

Name	Description
coral_expected_from_logits	Compute expected ordinal value from CORAL logits.
coral_logits_to_prob	Convert CORAL logits to a proper class distribution.
coral_loss	Compute CORAL loss (sum of binary cross-entropies over thresholds).
coral_random_loss	Expected CORAL loss for a random classifier.
build_root_eval_pointcloud	Build the root oracle evaluation point cloud for a rollout snippet.
canonical_fuse_points	Return finite points after deterministic voxel fusion and point capping.
observed_prefix_frame_indices	Return camera/depth frames observed at or before the rollout root time.
loss_key	Compose a logging key using the stage prefix.
metric_key	Compose a logging key using the stage prefix.
topk_accuracy_from_probs	Compute top-k accuracy from class probabilities.
chamfer_point_mesh	Compute accuracy, completeness, and bidirectional Chamfer for P<->M.
chamfer_point_mesh_batched	Chamfer-like point↔︎mesh distance per example (fully vectorised).
endpoint_log_gain	Compute endpoint log target-error gain.
endpoint_target_gain	Compute endpoint target-error gain J_e^(H).
finite_horizon_target_return	Compute additive selected target-RRI return G_t^(H).
selected_target_reward	Return the selected-step reward used for rollout/Q_H return.
selected_target_rri	Return the selected-step target RRI from one metric mapping.
summarize_target_rollout_metrics	Compute G_t^(H), endpoint gain, and log-gain for one trajectory.
target_point_mesh_error_after	Return selected-step target point-mesh error after adding the view.
target_point_mesh_error_before	Return selected-step target point-mesh error before adding the view.
ordinal_labels_to_levels	Convert ordinal labels to CORAL level targets.