search

RolandGao / RegSeg

The official implementation of "Rethink Dilated Convolution for Real-time Semantic Segmentation"
MIT License
86 stars 14 forks source link

Why not pretrain on ImageNet? #4

Open RobinhoodKi opened 2 years ago

RobinhoodKi commented 2 years ago

Hi, Thanks for your excellent work ! I notice that RegSeg can achieve a high accuracy on Cityscapes without pretraining. I also did a lot of ablation studies and I think DDRNet will drop around 3% miou if they do not use ImageNet pretraining. How about trying to train your encoder on ImageNet and see what will happen? I really look forward to your result ! Thanks !

RolandGao commented 2 years ago

Hi,

Sorry for not getting back to you earlier.

The large dilation rates in RegSeg do not work very well with small image sizes, so that's why I did not train on ImageNet. Recently, I'm trying to pretrain on another segmentation dataset called Mapillary and this should give a boost to the accuracy. I'll release the Mapillary pretrained models when I finish experimenting with them.