|
Hello I have a rather trivial question which I can not find the answer for. I am working with NNs and I am processing a large image where I need to do a scene parsing (labelling two classes of objects). To achieve this I am tilling the image to 15x15 pixel overlapping tiles, which are fed as an input to the NN. As an output during training I am providing the labels of just one pixel (the middle one) for each tile. My question is related to the percentage of the training. I will explain this with a simple example Lets say I have a 200x200 pixel image. Then for the training of the NN I take 10 random samples (pixels) over the image extent. For these 10 pixels I take a 15x15 tile around them. Finally I provide the label of just these 10 samples (so 10 labels only). Then I train the NN classifier with these data, finally I do the classification on the 200x200 image. Question: What is the percentage of the training data in relation to the total input data? Answer 1: (10) labels / (200x200) pixels = 0.00025 Answer 2 : (15 x 15) tiles * (10) labels (200x200) pixels = 0.5625 In this case which is the correct answer ? Thank you in advance |
