With SVM, adaboost or similar alogrithms, image training sets must be cropped with specific constraints (keep image cropping ratio the same, have object tightly cropped, same resolution)
In general, does a training set for a neural network based classifier need to be generated under the same constraints?
If not, which constraints can be removed or loosened?
In Classic Machine Learning the concept is the features are manually engineered.
Indeed, classic classifiers would require a pre defined input size.
Namely the input vector must have a pre defined size.
It might be pixels are features extracted (Preferably).
For Deep Learning we can make the input the actual image with minimal pre processing (Usually some kind of normalization and augmentation if needed).
Fully Convolutional Nets (FCN) can support arbitrary image size as input.
With some tricks, like AdaptiveAvgPool2d in PyTorch, one could achieve that even with some Dense / Linear / Fully Connected Layers.
