thucbx99

你好，我理解这里的问题是在有target domain上模型产生的伪标签情况下，直接用这个伪标签训练检测器效果如何。我们发现由于domain gap的存在，target domain上由检测器输出的伪标签是非常不准确的，论文中也汇报了这样做的结果（Table 6），相比于d_adapt有很大的性能差异。

这里只是为了实现上的方便，这样我们在开发算法的过程中可以实时监控d_adapt在目标域上的表现如何，例如做一些可视化case study。当然也可以把d_adapt实现为不需要目标域的标注就可以运行的形式，目前我们考虑大家可能主要还是会在Domain Adaptation的基准数据集上做实验或者改进，就保留了现在的实现

> Hi, it seems that their down link does not work. I download the weight from another project as a replacement. https://github.com/kazuto1011/deeplab-pytorch Maybe you can try it. Thanks for your...

Thanks for pointing this out. We are now also looking for links to this dataset (because the previous dataset was lost).

I think it may be because the pseudo labels are not accurate and lead to error accumulation. Can you provide the evaluation result (such as mAP) of each round?

请问是直接运行我们的代码吗，还是做了一些修改之后，看样子像是加了Pytorch Lightning之类的框架之后跑的

bbox adaptation的消融实验只需要在整个pipeline中移除这一步就可以了。category adaptation对应的baseline是用检测器自己的输出作为pseudo label，可以参考论文中提到的方法Unbiased Teacher，但我们发现在跨领域的情况下这种方法效果会较差。

是指Domain Adaptation Detection中Clipart的划分吗，可以直接下载我们提供的数据集，我印象中是直接使用全部的数据进行训练/测试，其中训练时不使用标签

We're sorry that the versions of torch and torchvision used in TLlib are somewhat out of date. At present, the issues that have received a lot of feedback are mainly...

Sorry for that. Recently, there are some problems with our cloud disk. To make matters worse, the GPU server where we backed up these datasets has recently been attacked and...