Open ilteralp opened 4 years ago
Seems like the Upsampling2D is the standard upscale technique across models in this repository. For example, in the Unet architecture, Upsampling2D is employed instead of Transposed convolution.
You may refer to the author's answer for a similar question here https://github.com/divamgupta/image-segmentation-keras/issues/17
Hi all, I was checking decoder implementation of SegNet and saw that Upsampling2D is used here. Keras provides bilinear and nearest upsampling for Upsampling2D according to the docs, so one of them must be used here. But SegNet decoder uses pooling indices computed in max-pooling, how are they provided here ? Thank you.