I am seeking for your advice with the topic related to segmentation.
Imagine the flying bird in the sky and a man taking a picture of that bird every second. There is very little change happening to the background and the actual bird with every picture/sample (some cloud movement and obviously bird wings/head moving as well).
Basically, I would like to train the network so that it will segment the bird on the N+1 sample based on the training made with prior N samples. Say, I have already segmented the bird on those N samples (i.e. created the mask of that bird), what would be the next step? What type of network/architecture is the most appropriate for such a task? How many samples I need in order to train that network given that I will boost my samples with data augmentation techniques? Particularly, I am looking for the method that takes less possible time to train (minutes rather than hours) and number of samples are also small enough (i.e. 5 samples).
Note that this example with the bird is just an abstract vision of a problem, the actual segmentation task can be related to any subject (for instance, medical CT scans of a patient), so that the model is trained only on those N samples and succesfully segment the object on N+1 sample - both training and prediction ideally should happen within few minutes time (is it possible)?
Just want to emphasize again that all N samples are very similar to each other, and N+1 sample is >90% similar to the previous N samples.
All suggestion are welcome, Thanks in advice.