there’s actually no fully connected layers in FOMO, it’s all 1x1 convolutions.
this diagram is mixing concept names e.g. the feature extraction is done with the MobileNet, and the heat map is calculated using a 1x1 Conv. i.e. these things aren’t different layers as such, so in a block diagram I think you should pick one set of names or the other.
what would be the difference between Non-car vs No car ?
See FOMO classifier capacity for details of the Keras model ( and note that the resizing and centroid calculation done outside of the keras model )