I was trying to implement the following paper: https://arxiv.org/abs/1610.01563 and I came across something that seemed ambiguous to me. On page 4, second paragraph, it says

After processing the image in VGG, the feature maps of a selection of layers (conv5_1, relu5_1, relu5_2, conv5_3, relu5_5; selected via random search) are rescaled and cropped to match an earlier layer (conv2_1 in our implementation). ....... Matching here means that we identify a pixel in the output of a convolution with the center of its receptive field in its input layer.

It is my understanding the model the paper uses (VGG), its intermediate layer's output was taken (for example conv5_1) and as its spatial dimension is less than the output of the earlier layers, it was upscaled to match the spatial dimension of an earlier layer's input conv2_1. I didn't clearly understand the method for upscaling. I also understand that the receptive field of a pixel means the area (region) of input image which produces the pixel as a output after a series of convolutions (and pooling).

If each individual pixel are upscaled via some upscaling algorithm, then my intuition is that as a result of the upscale, the output's spatial dimension will be much larger than that of conv2_1. (This is because the CNN filters convolve on overlapping regions). So then it might be needed to crop the upscaled feature map to match the spatial dimension of the conv2_1 layer. But, as this is ambiguous in the paper, I wanted to learn about what is the convention or the most popular way to do this? Any reference or point to resource would be helpful. Thank you.



You must log in to answer this question.