thanks for your great work and for sharing. I'm looking into using your Deeplab v3+ pre-trained model on Cityscapes, with MobileNet v2 backbone.
The model you shared (deeplabv3_plus_mobilenetv2_segmentron.pth) contains the full backbone but a somewhat light Deeplab v3+ head: no ASPP, no decoder. Is that because this is what gave you the best performance with the MobileNet v2 backbone?
If not, would you be able to share a full Deeplab v3+ head?
Hi,
thanks for your great work and for sharing. I'm looking into using your Deeplab v3+ pre-trained model on Cityscapes, with MobileNet v2 backbone.
The model you shared (
deeplabv3_plus_mobilenetv2_segmentron.pth
) contains the full backbone but a somewhat light Deeplab v3+ head: noASPP
, nodecoder
. Is that because this is what gave you the best performance with the MobileNet v2 backbone?If not, would you be able to share a full Deeplab v3+ head?
Thanks!