What a neuron looks at
The receptive field of a feature is the region of the original image that can influence it. Shallow features see a tiny patch, while deep features can see most of the image. Knowing the size tells you whether a layer can capture an object at all.
How it grows
Each layer expands the receptive field. Stacking small kernels grows it additively, while stride and pooling multiply the growth because each step now covers more original pixels.
- Kernel size adds its extent at every layer.
- Stride scales how fast the field grows downstream.
- Pooling behaves like a strided layer for this purpose.
A simple recurrence
The new receptive field equals the previous one plus the kernel minus one times the product of all earlier strides. The running stride product is the key bookkeeping term.
Why it guides design
If your target objects span a hundred pixels but your features only see forty, the network cannot integrate the whole object. Designers add depth, larger kernels, or dilation to grow the field to match the task.
Key idea
The receptive field grows additively with kernel size scaled by the running product of strides, and it must be large enough to cover the objects you want to detect.