placeframe-core 2.0.0.dev1364__tar.gz → 2.0.0.dev1370__tar.gz

{placeframe_core-2.0.0.dev1364 → placeframe_core-2.0.0.dev1370}/AGENTS.md

@@ -35,7 +35,7 @@ The package is flat: 16 leaf modules under `src/placeframe_core/`, no `__init__.
 
 **Confidence-calibration model.**
 
-- `calibration.py` — `SCHEMA_VERSION = 2`. `RawLocalizationMetrics` and `RawMapMetrics` are the per-query and per-map raw inputs. `Features.compute(localization, map_metrics)` produces a 10-feature vector: `log1p` of inlier / match / image / point counts, `reproj_error_median / image_diagonal`, plus four ratios and diversities. `ToleranceModel` carries `logistic_weights` (Features-shaped), `logistic_intercept`, and an isotonic table (`isotonic_x_breakpoints`, `isotonic_y_breakpoints`). `CalibrationArtifact` bundles a `tight` and `loose` `ToleranceModel`, `sigma_meas_alpha` / `sigma_meas_beta` for the measurement-covariance scaling `Sigma_meas = alpha * Sigma_pnp + beta * I_6`, `loose_min` / `tight_min` floors, and version metadata. `load_global_calibration(path, expected_pipeline_version)` raises `CalibrationLoadError` with a remediation message if the file is missing, the schema version mismatches, or the pipeline version mismatches. `apply_global_calibration(calibration, features)` returns `(tight, loose, True)` — the third element is always literal `True`.
+- `calibration.py` — `SCHEMA_VERSION = 3`. `RawLocalizationMetrics` and `RawMapMetrics` are the per-query and per-map raw inputs. `Features.compute(localization, map_metrics)` produces a 10-feature vector: `log1p` of inlier / match / image / point counts, `reproj_error_median / image_diagonal`, plus four ratios and diversities. `ToleranceModel` carries `logistic_weights` (Features-shaped), `logistic_intercept`, and an isotonic table (`isotonic_x_breakpoints`, `isotonic_y_breakpoints`). `CalibrationArtifact` bundles a `tight` and `loose` `ToleranceModel`, `sigma_meas_alpha` and the floors `sigma_meas_beta_rotation` (rad²) / `sigma_meas_beta_translation` (m²) for the measurement covariance `Sigma_meas = alpha * Sigma_pnp + diag(beta_rotation * I_3, beta_translation * I_3)` (`measurement_covariance_floor()`), `loose_min` / `tight_min` floors, and version metadata. `load_global_calibration(path, expected_pipeline_version)` raises `CalibrationLoadError` with a remediation message if the file is missing, the schema version mismatches, or the pipeline version mismatches. `apply_global_calibration(calibration, features)` returns `(tight, loose, True)` — the third element is always literal `True`.
 
 **Inference glue and typing shims (torch-aware).**
 

{placeframe_core-2.0.0.dev1364 → placeframe_core-2.0.0.dev1370}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: placeframe-core
-Version: 2.0.0.dev1364
+Version: 2.0.0.dev1370
 Summary: Placeframe domain logic: camera configs, coordinate transforms, metrics
 License-Expression: Apache-2.0
 Requires-Python: >=3.13

{placeframe_core-2.0.0.dev1364 → placeframe_core-2.0.0.dev1370}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "placeframe-core"
-version = "2.0.0.dev1364"
+version = "2.0.0.dev1370"
 description = "Placeframe domain logic: camera configs, coordinate transforms, metrics"
 readme = "README.md"
 license = "Apache-2.0"

{placeframe_core-2.0.0.dev1364 → placeframe_core-2.0.0.dev1370}/src/placeframe_core/calibration.py

@@ -4,11 +4,11 @@ import math
 import sys
 from pathlib import Path
 
-from numpy import interp, log1p
+from numpy import diag, dtype, float64, interp, log1p, ndarray
 from pydantic import BaseModel
 
 
-SCHEMA_VERSION = 2
+SCHEMA_VERSION = 3
 
 # Sentinel `pipeline_version` value that bypasses the pipeline-version check in
 # `load_global_calibration`. Use only with placeholder calibrations whose values
@@ -84,10 +84,18 @@ class CalibrationArtifact(BaseModel):
     tight: ToleranceModel
     loose: ToleranceModel
     sigma_meas_alpha: float
-    sigma_meas_beta: float
+    # Variance floors added to the measurement covariance's diagonal: rad² on the three rotation
+    # axes, m² on the three translation axes. Separate because one number can't be a sensible floor
+    # in both units.
+    sigma_meas_beta_rotation: float
+    sigma_meas_beta_translation: float
     loose_min: float
     tight_min: float
 
+    def measurement_covariance_floor(self) -> ndarray[tuple[int, int], dtype[float64]]:
+        """The 6x6 floor, ordered [rotation, translation] like the covariance it is added to."""
+        return diag([self.sigma_meas_beta_rotation] * 3 + [self.sigma_meas_beta_translation] * 3)
+
     def write(self, path: Path) -> None:
         path.write_text(self.model_dump_json(indent=2) + "\n", encoding="utf-8")