Ajai Chemmanam
Receptive Field of a CNN
Date: 2022-04-01
Written by Ajai Chemmanam

Receptive Field of a CNN

What is a receptive field?

𝗧𝗡𝗲 π—Ώπ—²π—°π—²π—½π˜π—Άπ˜ƒπ—² 𝗳𝗢𝗲𝗹𝗱 is the region on the input side of a layer on which it operates at a time to give output as features. A larger receptive field ensures that larger objects are detected.

How to add receptive field to a CNN?

  • Adding more Convolutional Layers can increase the receptive field linearly

  • Subsampling layers like pooling increase the receptive field multiplicatively

  • Sequentially placed dilated convolution increases the receptive field exponentially

π—˜π—³π—³π—²π—°π˜π—Άπ˜ƒπ—² π—₯π—²π—°π—²π—½π˜π—Άπ˜ƒπ—² 𝗳𝗢𝗲𝗹𝗱?

π—˜π—³π—³π—²π—°π˜π—Άπ˜ƒπ—² π—₯π—²π—°π—²π—½π˜π—Άπ˜ƒπ—² 𝗳𝗢𝗲𝗹𝗱 represents the region that contains input area with an impact on the output. In CNN's, the pixels at the center of a receptive field have a larger impact on the output.

  • Using Skip connections can reduce the effective receptive field.