Hi,it’s really a good work. However when I train it with voc 2 clipart, the results look like some wrong (about 20 map in my experiment which is not as good as other daod methods) . Can you tell me the reason? about sigma HOT 4 CLOSED

Hi, that is a good question.

We found that the default hyper-parameters and training schedules in our EPM baseline (fcos-based detector) are not feasible for Pascal 2 Other settings. I have spent lots of time tuning these adaptation settings. After conducting the following modifications, the results will be comparable with SOTA.

1. following TIA using both Pascal and style-transferred Pascal (07 and 12) as the training set.
2. setting a larger value for GRL_WEIGHT_P{3-7}, e.g., 0.2
3. pre-training the model with source-domain data until convergence (around 9w iterations).
4. reducing the learning rate to 0.0001 (it makes a big difference for fcos on Pascal VOC with an inappropriate learning rate)

We plan to release the reproducible configs for these extra benchmark settings in our extended version. (It seems that there are no existing fcos-based daod works reporting comparable results on these extra settings.) Nevertheless, since it has been a long time, I need some time to double-check the details. If you need the pre-trained model and configs in advance, you can leave an e-mail for me. Thanks.

from sigma.

I really need the pretrained model and configs about pascal 2 other settings and my qq email is [email protected]. thank you very much.

from sigma.

[email protected]

I have sent you through e-mail. Thanks for your attention to our work.

from sigma.

[email protected]

I have sent you through e-mail. Thanks for your attention to our work.

thank you very much!

from sigma.