3
$\begingroup$

The paper of Fully Convolutional Networks for Semantic Segmentation , gives the following image, enter image description here.

What do those numbers represent, 96, 256, 384, etc? Are them the feature dimensions? For instance, normal image input has 3, coming from RGB channels. In accordance with this figure, looks like the feature dimensions keeps growing along with the forward pass, are there any specific considerations for this kind of design?

$\endgroup$
0
$\begingroup$

Yes the numbers are feature dimensions. That's a common design for classification CNNs, decreasing the spatial dimensionality along the network to learn meaningful high-level features, while increasing the feature dimensionality to compensate for the loss of information.

You can think of low-level features as letters. There are in total 26 possible letters, so the feature dimensions should be small in the beginning.

High-level features are like words, which are combinations of low-level features (letters). In total there're thousands of commonly used words, so the feature dimensions should be large at the end.

The set of high-level features should not be equal to all possible combinations of low-level features, because a string like "meiluanyong" won't be very helpful with processing English texts.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.