search

PingoLH / FCHarDNet

Fully Convolutional HarDNet for Segmentation in Pytorch
MIT License
195 stars 52 forks source link

upsample problem #27

Open Tony-Hou opened 4 years ago

Tony-Hou commented 4 years ago

Can deconv replace bilinear upsample?

PingoLH commented 4 years ago

hi, sorry for the late reply. We did try deconv before, but as found in other research like ICNet, bilinear upsample is more efficient than deconv in terms of acc/latency. However, if deconv is as fast as bilinear interpolation on your deployment target, deconv can still be a good option.