Results are not good for dota dataset

[heiyuxiaokai]

The dataset is dota of whu
These images are cropped into (1200 x 1200) by step 900, about 8000 images.
1.The total mAP is about 0.38 when the global step is 80000, and some classes can not be detected, such as basketball_court, ground_track_field, soccer_ball_field
2.Total loss is declining from 8 to 0.4 when the global step is 260000(it seems to be the limit of this code?). But the mAP is 0.34 and some classes that can be detected in step 80000 can not be detected in step 260000.
I do some changes in the train.py for multi_gpus_train, and I have met the same problem when I use the origin code of one gpu.

[yangxue0827]

Recommend improved code: https://github.com/DetectionTeamUCAS/FPN_Tensorflow. @heiyuxiaokai

[WhiteSheep250]

@heiyuxiaokai
When you train the DOTA dataset, have you changed the parameters inside the cfgs.py? What are the specific changes? Thank you.

[heiyuxiaokai]

Reference there: DetectionTeamUCAS/FPN_Tensorflow#59. @WhiteSheep250

[WhiteSheep250]

Ok, I will try that method. Thank you. @heiyuxiaokai

[WhiteSheep250]

Does the new method fit the DOTA dataset?