The VGG16 classification network (CNN with fully connected layers) can be converted to a FCN (Fully Convolutional Network) by converting the fully connected layers to convolutional layers.
In addition, we can modify input layer to the VGG16 (which expects images of size 224x224x3) to accept any size (at least bigger than the size of the first conv layers filter) inputs.
With both of these modifications done, passing an input of say 448x448x3 will produce a predictions output of 14x14x1000. This is because the VGG16 network uses pooling layers which effectively scale down the input image by a factor of 32.
Each of the 14x14 2D submatrices of the 14x14x1000 predictions output corresponds to the a probabilistic heatmap of 1 of the 1000 output classes.
What is the corresponding input image coordinate for each of the entries in the 14x14 2D submatrices of the 14x14x1000 predictions output?
Naively, I'm guessing the  entry of the 14x14xi matrix corresponds to a probability of class i being located at input images [6.5][6.5] location (i.e. row 6.5, col 6.5) and the [