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

This 2D counterpart builds a real Feature Pyramid Network from scratch on canvas pixel data: a Sobel edge map feeds a five-level bottom-up backbone (C2–C6, block pooling with a blended average/max to stand in for growing CNN receptive fields), then a genuine top-down pathway nearest-neighbor-upsamples each coarser level and element-wise adds it to the finer backbone level's lateral projection, producing fused pyramid outputs P2–P6 that are simultaneously high-resolution and semantically stronger than the raw backbone. A real flood-fill connected-component detector finds the scattered objects directly from pixel data, and each detected object's actual bounding-box size drives the same scale-assignment formula used by Faster R-CNN-FPN and RetinaNet (k = round(4 + log2(√(wh)/224))), so you can watch — and verify live — that small objects land on fine pyramid levels and large objects land on coarse ones.