What is "classes" parameter in config? about mediar HOT 2 CLOSED

classes config is the model-specific config. It decides the number of final output neurons of the fully-connected layer.

As cell segmentation is binary problem (0:bgd, 1:cell), we either use 1 or 2 neurons. We used a single neuron for the segmentation.

In the config file, you can see we set 3 for the classes. Here, the other two dimensions each stand for x-axis and y-axis gradient field. They predicts the relative position w.r.t. center of the cell center (median of h,w).

from mediar.

Although we also mentioned in our paper as "Gradient Tracking" Sections, the act of how this tracking works is explained in more details at the CellPose paper. The gradient tracking parts in our code is a refactored version of their official code.

from mediar.