Deep Learning
Receptive Field
One convolution sees three pixels. Stack them and each layer sees three of the layer below, which saw three each of their own, so the window on the original image grows with depth. How fast it grows is decided by three numbers, and they do not all behave the same way.
the output cell you pickedeverything it can seepixels it will never see
Click any cell in the top row to trace a different output
Receptive field—
—
Dilation
Padding
receptive field—
distance between neighbours—
width of the top layer—
weights per layer—
What to observe
- With stride 1 and no dilation every layer adds k − 1 pixels, so the field grows only linearly with depth.
- A stride multiplies the reach of every later layer, and divides the width of every row above it: reach bought with resolution.
- Doubling dilation with same padding: the field grows exponentially, the weights per layer stay at k and the width never shrinks. The cone stays solid.
- Use dilation 2 in every layer and all offsets stay even, so only every other pixel is ever read. A shared factor in the dilations leaves holes in the field.
Shortcuts:f fullscreen