Feature Pyramid Network: Multi-Scale Object Detection

Modern object detectors don't run on a single feature map — they build a Feature Pyramid Network, a stack of feature maps at five resolutions (P2–P6) fused from a shrinking backbone via lateral 1×1 convolutions and a top-down upsample-and-add pathway, so every scale is both spatially precise and semantically strong. This simulator renders the full FPN topology in 3D: a backbone stack on the left, the fused pyramid on the right, and a placeable object whose pixel size and position drive the real scale-assignment rule (k = round(k0 + log2(√(wh)/224))) used by detectors like Faster R-CNN-FPN and RetinaNet to route each box to its matching level. Toggle the lateral connections and the top-down pathway independently to see why FPN needs both: drop either one and the level that lights up for your object loses either its fine detail or its semantic strength.