- To replace the use of
resize
orcrop
, I can use spp layer, i.e. spational pyramid pooling layer
or adaptiveAvgPool layer to have a fixed shape which can be used as the input of FC layer after the Conv layer. - But which will be preferred?
- I simply know that spp layer will extract input features using some different kernel shape into some fixed output shape and then flatten them
- adaptiveAvgPool can auto compute the kernel size and stride to get a fixed output shape