Differences between DIOR and VisdroneZSD datasets in training and testing horizontal box object detection tasks

[UnchangedOriInten]

What's the feature?

Hello author, I would like to ask about the difference between the DIOR and VisdroneZSD datasets in training and testing for horizontal box object detection tasks. Are both pre trained using softteachers? What are the labeled data, unlabeled data, and test data used during their two stages of training? Looking forward to your reply very much, thank you very much!

Any other context?

No response

[lizzy8587]

Hi @UnchangedOriInten, their share the same training configs. You can evaluate DIOR using this file dior_testset.txt.

[lili196]

Therefore, the training data of the Dior dataset and the visdronezsd dataset are the same, both use unlabeled data and labeled data, but the test data is different, so I want to train on the Dior dataset, do I only need to modify the path of the test data on the configuration file, is this the case?

[lili196]

Looking forward to your reply very much, thank you very much!

[lizzy8587]

Yes. They share the same training config & ckpt.

[lili196]

I still have some questions to ask, the paper uses A6000, but I don't have this card, can I use L20 to run, the batchszie in the code is 8, the training round is 80000, rpn_pseudo_thr=0.9,cls_pseudo_thr=0.9, but the experimental setup in the paper follows by 10k iterations training with a batch size of 12 on a single A6000 GPU. set the RPN foreground threshold and prediction
probability threshold p0 relatively high at 0.95 and 0.8, respectively. I'd like to reproduce whether I should use the configuration in the code or the configuration in the paper. We look forward to hearing from you, thank you very much

[lizzy8587]

Hi @lili196, I didn't test the code on L20, you can have a try. The config file is consistent with the paper: semi_rpn_score=0.95 and semi_cls_score=0.8 (used to filter pseudo boxes). The batch size is 12 in the code, i.e., batch_num=4 (in the pseudo queue) + batch_size=8.

[lili196]

when I run the python tools/train.py configs/visdrone_step1_base.py,
train_cfg = dict(
type='IterBasedTrainLoop', max_iters=80000, val_interval=10000),,Do I need to change 80000 to 10000?

[lizzy8587]

No, just keep the default setting.

[lili196]

Thank you very much

[wodexiaomalue]

Hi, @lili196 ,have you reproduced the results? looking forward your reply,thanks