search

jbwang1997 / OBBDetection

OBBDetection is an oriented object detection library, which is based on MMdetection.
Apache License 2.0
537 stars 113 forks source link

About the results on DOTA-v1.5 #47

Closed Richardych closed 2 years ago

Richardych commented 2 years ago

Hi,

Thanks for the great work, I wonder if you have done experiments on the DOTA-v1.5 which is a more challenging dataset.

Because our results with your Orient RCNN code on DOTA-v1.5 is only about 71% mAP.

jbwang1997 commented 2 years ago

We don't test our model on DOTA-1.5. But the baseline results (Faster RCNN + RT) achieves 65.03% mAP as we refer in AerialDetection model zoo. So I think 71% mAP is a pretty high performance if without augments.

Richardych commented 2 years ago

@jbwang1997 But I get the results with "configs/obb/oriented_rcnn/faster_rcnn_orpn_r50_fpn_1x_ms_rr_dota15.py" , which is multi-scale training with flip and rotation augments.

jbwang1997 commented 2 years ago

The multi-scale training needs to split images under scales 1.5 and 0.5. This operation is completed by BboxToolkit. I wonder if your skip this operation.

Besides, The results in AerialDetection model zoo are trained on the DOTA train and val set. Check if you only train the model on train set.

jbwang1997 commented 2 years ago

We have tested our method on DOTA-2.0 and the results are higher than Faster R-CNN+RT. So I think it's the same on DOTA-1.5.

Richardych commented 2 years ago

@jbwang1997 Can you tell us the results on DOTA-2.0 and the corresponding training and test config, Thanks.

jbwang1997 commented 2 years ago

Sorry for the late response. We only test the 2x schedule on DOTA2.0. here are the results. image