xuebinqin
commented
3 years ago Thanks for your interests. The only thing we changed in the TPAMI submission is the input resolution. (256x256 ->320x320). We didn't use any hyperparameters searching strategies or tools. Both BASNet and U2Net are trained with Adam. No tricks needed. If you follow the instructions described in the paper, you will get similar results.
Besides, BASNet and U2Net have different characteristics. BASNet are more likely to give you sharper boundaries, which are good at representing rigid targets. U2Net is good at capturing fine structures while giving relatively smoother transitions in the boundary regions, which is good to be used in segmenting soft targets, such as human hair or animal fur segmentation. In addition, U2Net weights are much smaller than BASNet. In my point of view, the evaluation metrics are just showing a general performance evaluation. Even some of the ground truth are not accurately labeled. Thus, higher "metric scores" doesn't necessarily mean "better performance" in different applications. It depends on your task and datasets.
On Tue, Apr 27, 2021 at 12:17 PM 张取 @.***> wrote:
您好！我拜读了您们提交在TPAMI 2021上面的最新版BASNet文章。与CVPR 2019相比，网络结果没有变化，不过在训练细节部分发生了改变。最后得到的模型通过指标评估，普遍超过2019年的模型，同时，我也对比了u2net，也在好几个数据集超过u2net. 请问您训练的超参数这些是通过自动搜索工具得到的，还是根据对比实验得出？比如 lr, batch size, iterations等等。 谢谢！
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/xuebinqin/BASNet/issues/55, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORNUDVIFHOMVT4GJG7TTKZXKHANCNFSM43UPUOTQ .
-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage:https://webdocs.cs.ualberta.ca/~xuebin/
您好！我拜读了您们提交在TPAMI 2021上面的最新版BASNet文章。与CVPR 2019相比，网络结果没有变化，不过在训练细节部分发生了改变。最后得到的模型通过指标评估，普遍超过2019年的模型，同时，我也对比了u2net，也在好几个数据集超过u2net. 请问您训练的超参数这些是通过自动搜索工具得到的，还是根据对比实验得出？比如 lr, batch size, iterations等等。 谢谢！