Jiaming Han

https://github.com/csuhan/ReDet/issues/39

1. Backbone得到的通道顺序是一致的，具有旋转等变性的，RiRoI Align通过通道的“对齐”，从旋转等变中得到旋转不变性。我们假设方向为θ的目标可以通过旋转变换T(θ)的得到。因此RiRoI Align相当于一个逆向变换T'(θ)，将目标的Region feature统一到0°，即rotation-invariant。 2. 是的。理论上我们可以使不同朝向的物体对齐到统一坐标中。 3. 不是。我们希望将C(r, n)调整到C(0, n)的位置。

Assume you have downloaded DOTA dataset, you can remove annotations of these classes by writing a class-filtering script (i.e., remove lines containing these classes in dota/labelTxt/*.txt). Then, you can use...

Try to set `workers_per_gpu=0` https://github.com/csuhan/ReDet/blob/0b9addf3c2734659fd6ffc7824f2e659fde4419c/configs/ReDet/ReDet_re50_refpn_1x_dota1.py#L143

目前oriented bbox是在数据的segmentation里编码的，即bbox仍是水平框，倾斜框由segmentation转成polygon。 因此，在data pipeline阶段，有针对seg增强的aug是有效的

参考多尺度数据生成：https://github.com/csuhan/ReDet/blob/master/DOTA_devkit/prepare_dota1_5_v2.py config：https://github.com/csuhan/ReDet/blob/master/configs/ReDet/ReDet_re50_refpn_1x_dota15_ms.py 测试结束后的result merge：python tools/parse_results.py --config configs/ReDet/ReDet_re50_refpn_1x_dota15_ms.py --type OBB

`ConcatDataset` may not support `show_result`. Use single dataset instead.

Copied from https://github.com/open-mmlab/mmclassification/blob/master/docs/getting_started.md > Download meta data using this https://github.com/BVLC/caffe/blob/master/data/ilsvrc12/get_ilsvrc_aux.sh

Yes. You can comment this line. However, BFP (with regualr conv) may break the rotation equivariance and invariance. You can try e2cnn-based BFP, just like our ReFPN.

I guess the reason is: cuda11.0 requires higher version pytorch (>1.3), while some [ops](https://github.com/csuhan/ReDet/tree/master/mmdet/ops) in our code are designed for pytorch