What does it mean of ori+type1?

[luoluo123123123123]

in Table 1, the WQT is used to create type1 data. Does ori+type1 mean that you put ori (4000 images) and also type1 (4000 images) total 8000 images of fully superviesed learning? Does it cause overfitting?

[mousecpn]

Actually, this is what the paper talks about. If you model is trained on one type and evaluated on another type, it will definitely cause overfitting. So, ori+type1 will overfit on type2,3,4,5,6,7,8. To alleviate the problem caused by domain gap (overfitting), the paper proposes DG-YOLO.

[luoluo123123123123]

thanks,but the wired thing is that ,ori+type1 can improve the mAP in ori. you said ori+type1 cause overfit on type2~8, but it still not bad on ori validation

[mousecpn]

It is obvious that ori+type1 can improve the mAP in ori. The training and the evaluation are in the same domain (both training data and test data contain ori). type1 increases the data diversity of the training data. So it can perform better on ori_val.

[luoluo123123123123]

I'm sorry, I try URPC2019 as ori, and type1=URPC2019+WQT (output scale = 600),and I use faster rcnn in detectron2, it can imporve the generalization ability in other types as the paper says, but when I test on URPC2019test (ori test),there is 2 mAP drop, I don't know if I set the wrong setting?Maybe the size should be 416?

[mousecpn]

Are the training size and the test size the same? If not, there will be a performance drop. Your WQT output size is 600, so you better not use a training size higher than 600.

[luoluo123123123123]

Oh,I forgot this point ,thanks a lot. May I ask two more questions in your paper? 1.As you said ,ori + type7 can improve 2mAP, an ori+type4 can also imporve 2 mAP , but why ori+type7+type4 do not improve 4mAP in ori? 2.Full WQT can imporve unseen type8 from 16 to 30, that means, for instance,we have a starfish in dark green,light green,light blue from WQT,but detector not only know starfish from those,but also generalize in type8, does it means WQT learns domain invariant information from dataset? If that is , why we still use DG-YOLO? You said DG-YOLO abandons the domain-related information tries to learn domain invariant information.That means,WQT just learn domain-related information,we put green starfish,and detectors can only learn green starfish,but WQT can generalize to unseen domain,and not drop in ori, why WQT can learn domain invariant information without abandons the domain-related information?

[mousecpn]

我猜你应该是一个中国人，我用中文回复好了，清晰一点。

1. 这种性能的提升不遵循线性关系。
2. 域泛化问题，本质上还是一个泛化问题（Generalization）。那个大表格的意思是想表达，我们测试域的性能和我们训练域与测试域之间的距离是有关系的，训练域和测试域越接近，那么测试域的性能就越高。但是，我们不可能能够知道未来在实际应用的时候，真正的测试域是什么样的（type8），而且仅仅是简单地把训练域的数量增加，也并不能完全解决这个问题（type8上30map的性能和同域训练测试的56+map性能还是有很大差距）。说明我们还需要一些别的技巧以让模型尽可能学习到域无关信息。当然DG-YOLO提出的模块确实比单纯用WQT带来了提升，然而这个结论也说了仅仅这样还是不够的，这个问题还需要进一步去研究。

同学可以参考我的最新工作，希望能够给你带来启发：https://github.com/mousecpn/DMC-Domain-Generalization-for-Underwater-Object-Detection