Does the implementation of SegNet use cached max pooling indices for fast upsampling as described in the paper? Looking at the implementation it seems that there is only a custom implementation of a decoder layer, but I don't see how it uses indices to upsample. It seems like it just learns the transpose convolution (deconvolution) layer just as a fully convolutional network would.
Does the implementation of SegNet use cached max pooling indices for fast upsampling as described in the paper? Looking at the implementation it seems that there is only a custom implementation of a decoder layer, but I don't see how it uses indices to upsample. It seems like it just learns the transpose convolution (deconvolution) layer just as a fully convolutional network would.
Is there something I'm missing?