Feature Pyramid or Hierarchical Model¶
Method — instantiates Cascaded Hierarchical Recognition
Represents input at multiple levels of resolution so broad structure and fine detail can be processed in order.
Feature Pyramid or Hierarchical Model is a representational method: it encodes the input at several resolutions at once — coarse global structure through fine local detail — as a stacked pyramid, so that later recognition can read the level appropriate to each question instead of scanning raw detail everywhere. Its defining idea is that it is the multi-scale representation the cascade reads — a data structure, not a gate, a workflow, or a screen. It supplies the levels; other mechanisms decide what to do at each. Build the pyramid once, and a broad question ("is there a structure of interest here?") is answered cheaply at the top while a fine question ("is it this class or that?") is answered only where the top said to look.
Example¶
An aerial-imagery system inspecting satellite photographs must find both large structures (runways, ships, storage tanks) and small ones (individual vehicles) in the same frame. Scanning full-resolution pixels everywhere for tiny objects is ruinously expensive; scanning only a downsampled image misses the small ones entirely. A feature pyramid resolves the tension by representing each image at several scales — an approach in the family of image pyramids and feature pyramid networks.[n1] Coarse levels cheaply flag broad structure and propose regions worth attention; fine levels are computed only inside those surviving regions, where they run the tests that separate a truck from a car. One representation serves both the context question and the detail question, in order.
How it works¶
- Build the multi-resolution representation. Downsample (and, in learned models, upsample and combine) the input into a stack of levels from coarse to fine.
- Attach a cue dictionary per level. Name which features are meaningful at each scale — global shape at the top, texture and fine edges at the bottom.
- Read broad first. Use coarse levels to locate structure and propose regions cheaply.
- Test fine only where it matters. In surviving regions, compute the fine-scale features that discriminate close candidates, rather than everywhere.
Tuning parameters¶
- Number of levels — more scales capture a wider range of object sizes but cost memory and compute.
- Downsampling ratio — how aggressively each level shrinks; coarser jumps are cheaper but can skip the scale a target lives at.
- Feature-to-level assignment — which cues are read at which scale; misassigning a cue reads it where it is meaningless.
- Shared vs level-specific detectors — sharing is efficient; specializing is more accurate per scale but heavier.
- Refinement stop — how deep into fine scales to go before deciding; deeper resolves subtle cases but costs.
When it helps, and when it misleads¶
Its strength is that a single pass yields both context and detail, and the cheap coarse levels prune the regions where fine detail is unnecessary — the compute goes where the discrimination is.
Its failure mode is reading a cue at the wrong scale: a small or subtle target can vanish at coarse resolution and never be looked for finely, a resolution-induced form of rare-case exclusion. The classic misuse is leaning on only the cheap top of the pyramid because it is fast, which systematically loses small and high-stakes cases. The guarding discipline is to ensure the finest scale is actually consulted for small or high-stakes targets rather than assumed away by a coarse pass.
How it implements the components¶
feature_hierarchy— the pyramid is the hierarchy: levels ordered coarse-to-fine by resolution, cost, and discriminative power.domain_feature_dictionary— each level carries a dictionary of the cues meaningful at that scale, from global shape to fine texture.discriminative_feature_test— the fine levels compute the tests that separate close candidates within a surviving region, rather than gathering detail indiscriminately.
It represents features but does not sequence classifiers or set the per-stage thresholds that actually reject candidates (stage_threshold_set, coarse_gate — that's multi_stage_classifier_pipeline); the pyramid is the map the cascade reads, not the gate that acts on it.
Related¶
- Instantiates: Cascaded Hierarchical Recognition — the multi-scale representation the cascade's stages read.
- Sibling mechanisms: Cascade Error Audit · Coarse Screening Checklist · Drilldown Recognition Interface · Expert Review Checkpoint · Multi-Stage Classifier Pipeline · Salience Map or Attention Heatmap · Stage Transition Log · Triage Queue with Escalation Rules
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Feature Pyramid or Hierarchical Model operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it represents input at multiple levels of resolution so broad structure and fine detail can be processed in order.
Independent corroboration: The frozen evidence defines Feature Pyramid or Hierarchical Model as 'Represents input at multiple levels of resolution so broad structure and fine detail can be processed in order', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Computer Science & Software Engineering
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Feature pyramids and multiscale hierarchical representations were developed in computer vision and machine learning.
Related originating lineages:
- Cognitive Science — Hierarchical perceptual processing materially shaped multiscale recognition architectures.
Review resolution: Both reviewers agree that computer_science is primary. I retain cognitive_science only as formative origin lineage(s), without treating every later application as an origin. cross_disciplinary_synthesis is appropriate because the exact artifact combines contributions from multiple professional lineages. Reach is specialized as a separate applicability judgment: it does not widen or narrow the recorded provenance. Encyclopedia synthesis is false because the artifact is already established enough that encyclopedia-specific synthesis is not required. The secondary differences are reconciled with no unresolved primary-provenance ambiguity.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] An image pyramid represents a signal at a set of successively coarser resolutions; a feature pyramid network is the learned analogue used in modern object detection, combining coarse semantic levels with fine spatial ones. Both make multi-scale structure available in a single representation so recognition can pick the right level per question. ↩