Interpret a spherical CAM

Level: Understand · Input: class-score lattice and explicit ERP support

A CAM is directional evidence: high values indicate directions that supported one classifier label. It is not an instance mask, a calibrated pixel probability or proof that the named object occupies every highlighted pixel.

Keep these distinctions explicit:

  • output stride controls CAM density; fully convolutional does not mean one score per input pixel;

  • resizing a CAM is visualization, not added semantic resolution;

  • ImageNet labels may act as fixed proxies for visually related industrial structures;

  • zero CAM is valid evidence; invalid or unobserved directions require a separate support mask;

  • changing ERP resolution changes angular tap spacing and receptive field.

Inspect the exact label, score, output lattice, solid-angle concentration, support and model provenance. Compare a spherical run with a controlled planar or tangent-view oracle before claiming equivalence.

Resolution and receptive field

For a stride-32 FCN, a $H\times W$ ERP produces approximately $\lceil H/32\rceil\times\lceil W/32\rceil$ logits, subject to the architecture’s padding rules. Every cell has a large receptive field. It says which broad directions influenced the class—not that a stride-sized image patch is an object box. Local-window aggregation can create denser evidence only when it is a declared model operation; interpolation cannot.

Interpret proxies conservatively

An ImageNet class can be a reproducible visual proxy for an industrial concept without being its literal identity. Keep both names: for example, the output hypothesis may say grid-like structure, supported by exact classifier label prison. If several proxies form one concept, freeze their non-negative weights before evaluation. Insufficient evidence yields unknown; it does not authorize a more convenient industrial label.

Useful summaries

Report peak direction, solid-angle-weighted mass inside a candidate region, concentration relative to the supported sphere, stability under longitude rotation and enrichment over unrelated control classes. A red overlay alone does not expose whether the maximum moved, whether support changed, or whether the class was globally relevant.

See the CAM tutorial for controlled scenarios. Next: form angular proposals.