Receptive Field & Parameter Explorer

Stack convolution layers, add pooling and dilation, and watch how quickly — or slowly — a network gains a global view of the image.

receptive field
CNN architecture
pooling
dilation

Final receptive field

18×18 px

Final spatial resolution

32×32

Total conv parameters

18816.000

Final receptive field: 18×18 pixels

Receptive field growth vs. depth

Architecture summary

Input size: 128 × 128 × 1

Layers: 3

Kernel: 3 × 3, dilation 1

Pooling: Yes (2×2)

Filters per layer: 32

Final spatial resolution: 32 × 32

Final receptive field: 18 × 18

Total conv parameters: 18,816

⚠ Receptive field does not cover the full input — the final layer cannot see the whole image at once.

Pooling buys receptive field, dilation buys it for free

Pure 3×3 convolutions add only 2 pixels of receptive field per layer — reaching a 512-pixel image would take roughly 256 stacked layers. Pooling (or strided convolution) halves spatial resolution each time, which doubles the effective receptive field of every subsequent layer: exponential growth instead of linear. Dilation inserts zeros into the kernel, expanding its footprint without adding parameters — a 3×3 kernel with dilation rate 2 sees a 5×5 region but still has 9 weights. Parameter count, meanwhile, grows with the product of input and output filters, independent of receptive field.

Preparing the simulation…