STDC-Seg icon indicating copy to clipboard operation
STDC-Seg copied to clipboard

Published 20 hours ago • verified publisher icon MichaelFan01

aux loss

Open TomMao23 opened this issue 3 years ago • 1 comments

Did you use the aux loss like BiSeNetV1 for the 1/32 and 1/16 feature maps during training? If so, are the weights 1,1,1 like BiSeNet?

TomMao23 avatar Jun 24 '21 13:06 TomMao23

one more question，in the paper “ The scale ranges in [0.125, 1.5] and cropped resolution is 1024 × 512 for training Cityscapes” does it mean 2048*0.125=256 as the shortest of image long side range(a little smaller than common)

TomMao23 avatar Jun 24 '21 14:06 TomMao23