accuracy on DOTA v2.0 dataset

[wjs0725]

Thanks for your great work. But when I use your config dotav2_test_dcfl_r50_1x.py to train the model on DOTA v2.0 dataset, I only achieve 28.86% map on test set.

this is the command used for training the model:
python -u tools/train.py /cluster/home3/wjs/mmrotate-dcfl/configs/dcfl/dotav2_test_dcfl_r50_1x.py --work-dir='./ckpt/dota2_4_dcfl'

and the log is shown below:
20231013_021328.log.json

After training, I generate the model's prediction on test set through:
python ./tools/test.py /cluster/home3/wjs/mmrotate-dcfl/configs/dcfl/dotav2_test_dcfl_r50_1x.py ./ckpt/dota2_4_dcfl/epoch_12.pth --format-only --eval-options submission_dir="./ckpt/dota2_4_dcfl/Task1_results"

then I submit the file to the online server, and this is the email I recieve:

I feel confused about the large gap on the result. Could you help me with this problem? Thank you very much!

[Chasel-Tsui]

Hi, thanks for your interest in our work. The training log looks good, and the validation performance looks good on the val set. Did you modify anything related to the config? and could you please check the image number of the test set to make sure that the full set set is used. Besides, maybe you can take a look at the generated submission files to verify whether the prediction results are in the correct format.