search

Dootmaan / DSRL

Implementation of CVPR 2020 Dual Super-Resolution Learning for Semantic Segmentation
MIT License
98 stars 17 forks source link

Regarding your final experiment setting #6

Open SuperStacie opened 3 years ago

SuperStacie commented 3 years ago

Hi there,

Thanks for sharing your codes. I wonder if the experience setting in your current train.py can achieve mIoU=0.6768 on Cityscapes. Like with deeplab v3 as SSSR backbone, init_lr=0.005 with fast decay speed and w_sr=w_fa=0.5? I see you set the default epoch for Cityscapes is 1000 and mentioned the model converges very fast. May I know how many epochs you trained to get a relatively good mIoU (~60%) (for me I trained for 10 epochs and the performance is around 30%). How about your experiment settings on VOC2012? I tried your bash script but the model didn't converge at all with in 50 epochs. Thank you in advance:)!

Dootmaan commented 3 years ago

Hi there,

Thanks for your question. I'm sorry that I can't remember too many details right now but I think the current setting is what I used when achieving an mIoU of 0.6768. On Cityscapes, if you use the original DeepLab v3+ you usually need to train the model around 400 epochs to achieve a moderate mIoU since the code contains many different kinds of augmentations. However, after adding SISR and FA the model only needs about 150 epochs to achieve a similar mIoU.

I'll upload the pretrained weights soon.