Feature Pyramid Network: Multi-Scale Object Detection (2D)

Scene

Object count 5

Max object size 200 px

Network Pathways

Detection & Scale Assignment

Objects detected

–

Levels used

–

Scale-match check (small→fine, large→coarse)

–

#	size (px)	level	stride

How it works

This is a from-scratch, real (simplified) Feature Pyramid Network running entirely on 2D canvas pixels — no pretrained weights, but every step is a genuine computation on the image, not a scripted illustration.

1. Sobel edge magnitude over the raw 256x256 image  (pixel-level feature)
2. Bottom-up backbone (block pooling, avg+max):
     C2 64x64  ->  C3 32x32  ->  C4 16x16  ->  C5 8x8  ->  C6 4x4
   (each stage halves resolution, keeps the strongest local activation
    — a stand-in for a real CNN's growing receptive field / semantics)
3. Top-down pathway + lateral connections (coarsest to finest):
     P6 = C6
     P_i = Upsample2x_NN(P_{i+1}) + C_i        for i = 5,4,3,2
4. Detection: threshold vs. background + flood-fill connected components
   on the real pixel data -> genuine bounding boxes and sizes, not
   hand-placed coordinates.
5. Scale assignment (same rule real detectors use, Lin et al. 2017):
     k = round(4 + log2(sqrt(w*h) / 224)),  k clamped to [2, 6]

Backbone row (C2…C6) — bottom-up features, before fusion: coarse levels lose spatial detail.
Pyramid row (P2…P6) — after the top-down pathway: every level now carries both fine detail and coarse-level context.
Lateral connections OFF — P_i drops the C_i term, so fine spatial detail from the backbone never reaches the pyramid.
Top-down pathway OFF — P_i = C_i (raw backbone only), so coarse-level semantic context never flows down to fine levels.
The table's assignment is computed live from the actual detected bounding-box size of each object — not looked up from a fixed list.

Synthetic scene — detected objects boxed & colored by assigned pyramid level

Backbone C2–C6 (bottom-up, before fusion) Pyramid P2–P6 (after top-down fusion)

Backbone (bottom-up)

Pyramid (top-down fused)