search

chaoqichen / HTCN

Implementation of "Harmonizing Transferability and Discriminability for Adapting Object Detectors" (CVPR 2020)
111 stars 23 forks source link

How did you use CycleGAN generated pascal images for adaptation from pascal to clipart? #13

Open TADanton opened 3 years ago

TADanton commented 3 years ago

Dear chaoqi,

first of all, thanks for your nice work! I found the paper quite interesting and would like to run your code to see results.

In your paper, it is mentioned that pascal->clipart experiments follow "Strong weak distribution alignment" paper. Supplementary materials of "strong-weak" paper says CycleGAN "Translated images are utilized just for training detection modules and not utilized for domain classification."

May I ask how this usage is reflected in your code? Or how should I use your code to achieve this? Currently it seems I can only specify one source domain, which is either original pascal or transferred pascal (in the context of pascal->clipart). And the code seems to use all images from source domain for domain classification.

Directly using your code to adapt from original pascal to clipart reaches 0.3595 mAP, only using transferred pascal reaches 0.3686 mAP. Both are quite far away from reported 40.3 mAP in your paper. So I guess I didn't use the datasets correctly.

Could you please offer me some help? Thank a lot in advance!

Best regards, Anton

chaoqichen commented 3 years ago

@TADanton Hi! It seems that you may misunderstand my statement. We follow SWDA in the dataset setting instead of the technical details. And our code on pascal voc --> clipart can achieve 40.5 mAP even without using the translated images. Please make sure you have downloaded our latest code and check experimental settings including the weights of different losses (cf. our paper and supplementary materials). BTW, I can reply to you in detail after the CVPR DDL.

TADanton commented 3 years ago

Thanks for your quick reply!

I used your latest original code, so the weights of losses should be coorect?

Did you train and evaluate following "strong weak" paper, where they use clipart train+test during training without annotations and use clipart train+test again for evaluation?

In evaluation, your code reports an error asking for "val" set of clipart, which clipart doesn't have. I stacked train+test together to create a val.txt. Though, evaluating using train+test or only test doesn make a big difference in mAP, both are quite far away from reported 40.3mAP in your paper...

Besides, supplementary material of this paper doesn't seem to be published in internet, would that be possible to share it here? Or directly to my Email box? TAD_anton@hotmail.com

Thanks and wish you a successful submission to CVPR2021 !

kailigo commented 3 years ago

Same here. I am unable to reproduce the results. Without making any change on the code, I can get only 38.2 for the mAP for the adaptation from Pascal VOC to Cilpart, while the reported results is 40.3. Could you help figure out the reason? Thanks.