The concept of Superpixel Pooling Layer can be found in the paper "Weakly Supervised Semantic Segmentation Using Superpixel Pooling Network".

The general idea of superpixel pooling is very similar to regular pooling. Given a feature map of the image (output of the Neural Network) and a superpixel segmentation (eg. Felzenszwalb segmentation), the features over a superpixel are aggregated (per channel) by means of a reduction function (e.g. max() or avg()). This results in a matrix of size KxC in which C is the number of channels of the feature map and K is the number of superpixels (one feature per channel and per superpixel). Then, this superpixel-feature-map matrix is averaged over the superpixel dimension, obtaining a single feature vector of size 1xC, for the input image.

I do not understand how can you "combine" this superpixel-feature-vector of dimension 1xC with the actual feature map of the image (of size HxWxC) for its posterior segmentation. Is it obtained a "segmentation-score" per superpixel? Or is this superpixel-feature-vector combined with the feature map of the image in another manner for its posterior segmentation?

  • $\begingroup$ In this paper they make a superpixel classifier that takes as input the superpixel features extracted with the Superpixel Pooling Layer. $\endgroup$ – IgnacioGaBo Mar 8 at 14:23

Your Answer

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

Browse other questions tagged or ask your own question.