0
$\begingroup$

"Feature Pyramid Network" is a network that is used for feature extraction. Since it is pyramid in shape, it might be called so.

Consider the following excerpts from two different sources

#1

A Feature Pyramid Network, or FPN, is a feature extractor that takes a single-scale image of an arbitrary size as input, and outputs proportionally sized feature maps at multiple levels, in a fully convolutional fashion. This process is independent of the backbone convolutional architectures. It, therefore, acts as a generic solution for building feature pyramids inside deep convolutional networks to be used in tasks like object detection.

#2

FPN composes of a bottom-up and a top-down pathway. The bottom-up pathway is the usual convolutional network for feature extraction. As we go up, the spatial resolution decreases. With more high-level structures detected, the semantic value for each layer increases.

Both the sources are giving a strong sense that the phrase "Feature Pyramid Network" must be used for CNN's only as it is used mainly intended for object detection. But the name and purpose suggest to me that any ANN that is pyramid in shape can be attributed as "Feature Pyramid Network" since any ANN tries to extract features only in the general sense.

Am I true? Can I use the phrase for any arbitrary ANN that is pyramid in shape or is it an exclusive term of CNNs in computer vision only?

$\endgroup$

1 Answer 1

1
$\begingroup$

No, Feature Pyramid Networks does not only refer to CNNs.

Take the recently trending "Swin Transformer" as an example.

enter image description here

Swin Transformer is a variant of vision transformers, and they do not employ convolutional layers. However, their design allows them to output feature maps in multiple scales (feature pyramids). The authors actually used the feature pyramids for image segmentation tasks using UperNet.

The reason why people use feature pyramids is to use both local and global representations (mostly in images), and CNNs have a structure that naturally outputs such feature pyramids. However, this does not mean that feature pyramids are constrained to only CNNs.

$\endgroup$
5
  • $\begingroup$ I am not an expert on this topic, but I heard of "Feature Pyramid Networks". Do the authors in the Swin Transformer call their approach "Feature Pyramid Networks". Maybe what you mean is that similar approaches to "Feature Pyramid Networks" have also been developed with other layers? Because the "original" FPNs seem to have been proposed here, so it's possible that some people call FPNs only those that are consistent with that paper, or the term may have started to be used more loosely for similar ideas. $\endgroup$
    – nbro
    Jan 6 at 9:34
  • $\begingroup$ If I remember correctly, related ideas were already developed a long time ago, with SIFT. Again, I am not an expert in FPN, so it's possible I am making here some wrong assumptions. $\endgroup$
    – nbro
    Jan 6 at 9:35
  • $\begingroup$ I agree that FPN is a loosely used term, but the main idea is to aggregate hierarchical feature maps regardless of the backbone. In Swin Transformers, the authors used Swin Transformers as the backbone to extract their hierarchical feature maps. I found a quote from the paper "With these hierarchical feature maps, the Swin Transformer model can conveniently leverage advanced techniques for dense prediction such as feature pyramid networks (FPN) [42] or U-Net [51]." $\endgroup$
    – DKDK
    Jan 6 at 9:53
  • $\begingroup$ In that sentence, it seems that they are saying that their approach can use/leverage FPNs, so this suggests that they not the same thing. $\endgroup$
    – nbro
    Jan 6 at 9:54
  • $\begingroup$ hanugm asked whether the phrase "Feature Pyramid Network" must be used only for CNN's or not. My answer is that it is not only applicable for CNNs but any networks (like swin) that outputs hierarchical feature maps. $\endgroup$
    – DKDK
    Jan 7 at 1:19

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .