Segmentation model and backbone

[liheyoung]

Segmentation model and backbone

[LiheYoung]

The ResNet backbone in ST++ uses a 7x7 conv stem, while in UniMatch, we follow recent works to use a deep stem of three convs. And the output stride in our UniMatch for the Pascal is 16, instead of 8 in ST++.

[firelight8752]

Thanks.
Recent work refers to U2PL? Is the segmentation network structure of unimatch and U2PL identical?
In addition, what is the output stride in UniMatch for the Cityscapes?

[LiheYoung]

In the arXiv v1 (we may update it in the future), on the Cityscapes, our UniMatch follows U2PL to adopt the output stride of 8. The dimensions in U2PL decoder are also different from most works, you may refer to Haochen-Wang409/U2PL#25. Our decoder is the same as that in CPS and ST++.

[firelight8752]

Thank you for your reply. I understand these details.
It seems that these details are not mentioned in most papers. Can they be regarded as fair comparison?

[LiheYoung]

In my experience, the minor modifications in U2PL do not necessarily bring improvement. You may choose them according to your preference.

[firelight8752]

Thanks a lot.