Evaluate semantic regions and hypotheses

Level: Evaluate · Evidence: directional, regional and cross-view

Evaluate each stage separately:

  1. Classifier/FCN: planar parity on the canonical crop and complete layer replacement.

  2. Spherical CAM: seam continuity, supported solid-angle concentration and comparison with controlled tangent-view evidence.

  3. Proposal: object coverage, excess FOV, centre stability and duplicate suppression.

  4. Association: transferred-region overlap, track support, appearance consistency and agreement across views.

  5. Spatial hypothesis: reprojection residual, number of independent views, position/scale uncertainty and rejection state.

  6. Optional mask: spherical IoU and continuity only when a legitimate reference mask exists.

Do not report CAM as segmentation accuracy or use a selected attractive overlay as the evaluation. Fix proxy vocabularies and thresholds before the evaluation set, include failed channels and retain model, checkpoint, support, pose and proposal provenance.

Stage-specific measures

Stage

Useful measures

Misleading substitute

ported classifier

global-logit cosine, top-k retention, planar-layer audit

attractive CAM colour

CAM

angular peak error, solid-angle enrichment, rotation stability

upsampled pixel sharpness

proposal

target coverage, excess solid angle, centre/FOV stability, recall

centre-point hit alone

association

transferred IoU, track support, embedding agreement, ID precision/recall

nearest CAM peak

spatial estimate

angular reprojection, view count, uncertainty, metric error when scale exists

arbitrary-scale distance

optional mask

solid-angle IoU, boundary continuity, alternative-mask spread

planar ERP pixel IoU alone

Report proposal recall together with proposal count and area. A method that covers every target by proposing most of the sphere is not useful. For association, split true repeated instances across views so that semantic label agreement alone cannot solve the task.

Minimum reproducibility record

Freeze images and splits, proxy labels and weights, checkpoint hashes, ERP resolution, CAM layer/stride, support masks, proposal thresholds, anchor scales/FOVs, pose source, matching configuration and random seeds. Publish negative cases and all compared variants under the same visualization scale.

Use benchmarking for run records, the CAM tutorial for the CAM protocol and the object-hypothesis tutorial for association metrics.