feat(perception): restore camera-first semantic candidate

2026-07-30 21:53:34 +03:00
parent fdcaf57ae7
commit e4e5bf6ea1
6 changed files with 1436 additions and 0 deletions
@@ -0,0 +1,164 @@
+from __future__ import annotations
+
+from typing import Any
+
+import numpy as np
+
+from k1link.compute.e52_camera_first_restoration import (
+    E52_CASE_SCHEMA,
+    evaluate_camera_first_case,
+)
+
+PROFILE: dict[str, Any] = {
+    "target_labels": ["person", "car", "bus", "truck"],
+    "label_groups": {
+        "person": ["person"],
+        "vehicle": ["car", "bus", "truck"],
+    },
+    "minimum_detector_score": 0.5,
+    "maximum_bbox_image_fraction": 0.25,
+    "minimum_selected_lidar_points": 4,
+    "minimum_selected_lidar_coverage": 0.7,
+    "minimum_original_bbox_iou": 0.25,
+}
+
+
+def _item(*, stratum: str, snapshot: dict[str, Any]) -> dict[str, Any]:
+    return {
+        "item_id": "e30-review-item-" + "a" * 64,
+        "stratum": stratum,
+        "evidence_binding": {"source_frame_index": 259},
+        "materialization": {
+            "projection_width": 800,
+            "projection_height": 600,
+        },
+        "e29_snapshot": snapshot,
+    }
+
+
+def _case(*, stratum: str) -> dict[str, Any]:
+    return {
+        "sequence": 7,
+        "source_stratum": stratum,
+        "prediction": {
+            "presence": "occupied-environment",
+            "geometry_association": "independent-occupied",
+            "freshness": "current",
+        },
+        "reference": {
+            "presence": "object-present",
+            "geometry_association": "object-associated",
+            "freshness": "current",
+        },
+    }
+
+
+def _frame(instances: list[dict[str, Any]]) -> dict[str, Any]:
+    return {
+        "schema_version": "missioncore.panoptic-frame/v1",
+        "frame_index": 259,
+        "instances": instances,
+    }
+
+
+def test_e52_restores_camera_semantic_then_derives_lidar_range() -> None:
+    pixels = np.asarray([[150.0, 225.0], [152.0, 230.0], [155.0, 235.0], [160.0, 240.0]])
+    depths = np.asarray([4.0, 4.2, 4.4, 4.6])
+    result = evaluate_camera_first_case(
+        item=_item(
+            stratum="geometry-only",
+            snapshot={
+                "geometry_status": "single-source-geometry",
+                "nearest_range_m": 3.9,
+            },
+        ),
+        e40_case=_case(stratum="geometry-only"),
+        detector_frame=_frame(
+            [
+                {
+                    "instance_id": 1,
+                    "label": "person",
+                    "score": 0.99,
+                    "box_xyxy": [140.0, 210.0, 220.0, 450.0],
+                }
+            ]
+        ),
+        projected_pixels_xy=pixels,
+        projected_depth_m=depths,
+        projected_selected_mask=np.ones(4, dtype=np.uint8),
+        profile=PROFILE,
+    )
+
+    assert result["schema_version"] == E52_CASE_SCHEMA
+    assert result["restoration"] == {
+        "state": "camera-semantic-restored-with-lidar-range",
+        "semantic_owner": "camera",
+        "semantic_label": "person",
+        "metric_geometry_owner": "lidar",
+        "range_m": 4.3,
+        "runtime_publishable": False,
+        "requires_shadow_acceptance": True,
+    }
+
+
+def test_e52_rejects_pathological_full_frame_detector_box() -> None:
+    pixels = np.asarray([[150.0, 225.0], [152.0, 230.0], [155.0, 235.0], [160.0, 240.0]])
+    result = evaluate_camera_first_case(
+        item=_item(
+            stratum="geometry-only",
+            snapshot={"geometry_status": "single-source-geometry"},
+        ),
+        e40_case=_case(stratum="geometry-only"),
+        detector_frame=_frame(
+            [
+                {
+                    "instance_id": 1,
+                    "label": "car",
+                    "score": 0.9,
+                    "box_xyxy": [0.0, 0.0, 800.0, 600.0],
+                }
+            ]
+        ),
+        projected_pixels_xy=pixels,
+        projected_depth_m=np.asarray([4.0, 4.2, 4.4, 4.6]),
+        projected_selected_mask=np.ones(4, dtype=np.uint8),
+        profile=PROFILE,
+    )
+
+    assert result["candidate_camera"]["selected"] is None
+    assert result["candidate_camera"]["rejected_pathological_large_box_count"] == 1
+    assert result["restoration"]["state"] == "unresolved-no-camera-semantic"
+
+
+def test_e52_corroborates_existing_camera_semantic_without_inventing_range() -> None:
+    result = evaluate_camera_first_case(
+        item=_item(
+            stratum="camera-only",
+            snapshot={
+                "label": "bus",
+                "score": 0.4,
+                "bbox_xyxy": [100.0, 100.0, 200.0, 200.0],
+                "geometry_status": "camera-only",
+                "range_m": None,
+            },
+        ),
+        e40_case=_case(stratum="camera-only"),
+        detector_frame=_frame(
+            [
+                {
+                    "instance_id": 4,
+                    "label": "truck",
+                    "score": 0.95,
+                    "box_xyxy": [105.0, 105.0, 205.0, 205.0],
+                }
+            ]
+        ),
+        projected_pixels_xy=np.empty((0, 2), dtype=np.float64),
+        projected_depth_m=np.empty(0, dtype=np.float64),
+        projected_selected_mask=np.empty(0, dtype=np.uint8),
+        profile=PROFILE,
+    )
+
+    assert result["restoration"]["state"] == "camera-semantic-corroborated-no-lidar-range"
+    assert result["restoration"]["semantic_label"] == "truck"
+    assert result["restoration"]["range_m"] is None