Convert CAM evidence into angular proposals

Level: Compose · Output: finite spherical centre and FOV candidates

The useful analogy to the R-CNN family is proposal handling, not rectangular ERP boxes:

Detector pipeline	Spherical semantic pipeline
objectness map	CAM/proxy evidence
image box proposal	angular centre + horizontal/vertical FOV
box IoU/NMS	solid-angle overlap/NMS
ROI crop	gnomonic view centred on the proposal
box refinement	centre and FOV refinement

A proposal should cover enough of the candidate object for a local gnomonic view without becoming an entire-room hypothesis. Generate several centre/FOV anchors around stable CAM regions, score their supported evidence, reject unsupported area and suppress near-duplicates with solid-angle overlap.

From score map to candidate components

semantic_regions_from_map() provides the first deterministic split. It thresholds a normalized score map, joins components with horizontal seam wraparound, associates supported spherical features, and reports rejected, truncated and seam-crossing counts.

from panorai.object_localization import (
    SemanticRegionProposalConfig,
    semantic_regions_from_map,
)

components = semantic_regions_from_map(
    cam_hw,
    erp_shape_hw=erp_rgb.shape[:2],
    features=features,
    region_id_prefix="view-a/grid",
    class_id=class_id,
    class_name=class_name,
    config=SemanticRegionProposalConfig(
        threshold=0.50,
        min_component_cells=2,
        min_feature_count=2,
    ),
    source_id="resnet18-imagenet/cam",
)

The threshold is an experimental proposal parameter, not a calibrated probability boundary. A component without enough accepted features remains in the diagnostics but cannot support later geometric association.

Anchor centre and extent

For each retained component, generate centre/FOV anchors at several angular scales and aspect ratios. Score an anchor by supported semantic mass, feature/track support and penalties for unsupported or excessive area. Use minimum and maximum FOV bounds; otherwise a peak becomes a tiny crop or a weak component expands to the whole room.

Compare anchors with solid-angle intersection-over-union. The merge rule must distinguish duplicate observations of one region from incompatible concepts. This is the spherical counterpart of anchor generation, box regression and NMS—not an ERP bounding-box approximation.

Each proposal retains:

concept and exact proxy labels
centre bearing
horizontal and vertical FOV
semantic support score
support/validity coverage
source model, CAM and view provenance

The proposal is intentionally uncertain. Cross-view evidence decides whether it should be rejected, merged, kept ambiguous or promoted.

Next: associate proposals across views. The full method and ablations are in the object-hypothesis tutorial.