Jiaming Han

@liuyanyi I also notice the performance gap in my re-implemented mmdet_v2 (s2anet is first implemented with mmdet_v1). To align with detectron2, mmdet_v2 changes some lr&optimizer params. One possible solution is...

由于HRSC2016 a)不需要split image&merge results，b)没有HBB task。请将下面代码注释掉： https://github.com/csuhan/ReDet/blob/0b9addf3c2734659fd6ffc7824f2e659fde4419c/tools/parse_results.py#L96-L120

这里**注释掉一部分**指的是HRSC2016测试，DOTA不需要注释。

https://github.com/csuhan/ReDet/issues/15

由于计算资源有限，并没有尝试训练ReR101.

Similar to DOTA, please change the `config`: `python tools/parse_results.py --config configs/ReDet/ReDet_re50_refpn_1x_dota15.py --type OBB`

请查看是否生成结果，忽略错误提示

These methods may work well on some datasets, while for other datasets, adjusting highper-params are neccesary. I sugguest you to deeply analyse why these trick may help.

Sure. Please check our [ReDet_mmcls](https://github.com/csuhan/ReDet/tree/ReDet_mmcls) branch.

请确保单卡learning rate设置正确