Localize and associate semantic regions

Outcome: turn directional learned evidence into auditable angular regions and, when multiple registered views agree, spatial object hypotheses.

Semantic directional evidence in an industrial panorama

pretrained CNN -> spherical FCN -> CAM -> angular proposal
                                         |
                      registered views --+-> spatial hypothesis
                                         |
                                         +-> optional mask refinement

Start

  1. Generate a spherical CAM with a pretrained classifier.

  2. Interpret its resolution and support without treating it as a segmentation mask.

Build proposals

  1. Convert evidence into finite angular and FOV proposals.

Associate views

  1. Associate proposals across registered views and create a spatial hypothesis only when the evidence supports it.

Extend and evaluate

  1. Optionally refine an accepted region with local masks.

  2. Evaluate every stage without claiming ground-truth segmentation.