why the b0 results is so bad at cityscapes, when I train from scratch(don't use pre-trained backbone)

NVlabs / SegFormer

Official PyTorch implementation of SegFormer

https://arxiv.org/abs/2105.15203

Other

2.59k stars 357 forks source link

why the b0 results is so bad at cityscapes, when I train from scratch(don't use pre-trained backbone) #19

Closed GitHubChrischen closed 3 years ago

xieenze commented 3 years ago

(1) Pretrain is essential for semantic segmentation. (2) Cityscapes dataset is very small with only 2000 training images. (3) Other CNN network without pre-train also performs bad.

puyiwen commented 2 years ago

(1) 预训练对于语义分割是必不可少的。 (2) Cityscapes 数据集非常小，只有 2000 张训练图像。 (3) 其他没有预训练的 CNN 网络也表现不佳。

hi,How will I train mit_b4 pre-trained model?