amingwu/vdd-daod

Vector-Decomposed Disentanglement for Domain-Invariant Object Detection

vdd-daod's Issues

大佬，im_cls_lb这个是代表什么意思？

大佬，我看训练文件中有这句代码：im_cls_lb.resize_(data_s[2].size()).copy_(data_s[2])，im_cls_lb是表示什么意思，我有点不太明白，希望能解答一下，谢谢

How to get heatmap?

Thanks for your great work. I have a question that how to get the heatmap posted on Figure 8 in your paper. Could you give me some advice, maybe it's a related repo or a useful method. Thanks!

Import Error:_C.cpython-36m-x86_64-linux-gnu.so

When I run CUDA_VISIBLE_DEVICES=4 python3.6 da_train_net_gl.py --max_epochs 20 --cuda --dataset dc --dataset_t nr --bs 1 --da_use_contex --lc --gc, I meet the following error
ImportError: /home/lbb/DLProjects/VDD-DAOD/lib/model/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN6caffe28TypeMeta21_typeMetaDataInstanceIdEEPKNS_6detail12TypeMetaDataEv
my environment is:
python 3.6, pytorch 1.7.1, cuda 11.0

cannot import name '_C' from 'model'

Thank you to share your amazing code!

I want to run your training code, but when i run da_train_net_gl.py
I meet follow message: ImportError: cannot import name '_C' from 'model'

I unzip lib/model.zip, lib/roi_da_data_layer.zip lib/roi_data_layer.zip to the project.
And my environment is:
python 3.7 pytorch 1.7.1 cuda 11.2

Trend of domain classification loss?

Thanks for releasing the code of your interesting research.
I'm experimenting with your method on different datasets in different tasks.
As a result of the experiment, the domain classification loss in the source domain increases and the domain classification loss in the target domain gradually decreases. (Total domain classification loss decreases.) Could you tell me the correct trend of the domain classification objective function?
Looking forward to your reply. Thank you.

Why do you use GRL layer?

Thanks for your interesting work. I'm impressed with your method which is simple and effective.
I'm looking forward to your code release.

Also, I have a question about your method.

In your method, the GRL (Gradient Reversal Layer) is used in the domain classifier to reverse the gradient with respect to the domain-specific features. However, I don't understand why you add GRL in the domain classifier. In order to make domain-specific features be more domain-specific, I think that domain classification loss should be minimized, not maximized.

If there is any point I misunderstood, please let me know.

Thank you.

How can you evaluate the performance of the bdd's three new datasets?

Thank you for your code. However, when I download the datasets from your shared link, I found that there's no "val.txt". As a result, what's your approach to evaluate these datasets?

大佬，update文件夹里的txt文件是什么意思

Unable to reproduce the results

Hi, thank for you work and open-source code. I evaluated the Night-Rainy dataset using the provided checkpoint and results are matching with Table 4 (i.e 23.0 mAP) but when I trained model (Daytime-sunny → Night-rainy) on my local machine, I am only able to get 17.9 mAP. Is there anything I am missing during the training? Also how much iteration it took to train the provided checkpoint for Daytime-sunny → Night-rainy.
Thanks.

About the dataset Daytime-sunny, Night-rainy, Dusk-rainy

Hi, thanks for your work, I am wondering if you would release the annotations and corresponding image ids of these three datasets? Thx!

Code Environment

Thank you for your code release and management. I'd like to run your code.
Could you let me know the environment of your setting?
such as python, torch version, cuda, whether detectron or faster-rcnn, ...

The implementation of vgg16

Hi，Thanks for your great work.
It seems that the implementation of vgg16 is not complete?