Panxjia comments

Results 17 comments of


                                            Panxjia

Questions on benchmark

@JrPeng Hi, any progress about training the model with BN-backbone and gn-FPN/heads without using PA-structure ? Thanks!

question about conf_thresh

Sorry for the late reply. We set the conf_thresh to 0.05.

python rotate_augment.py path/to/images?

Sorry for the late reply. I have reported the problem in an earlier issue. We also encountered the problem because of the incomplete data provided by the SKU110K author. However,...

Centrernet training dota data set

我们是在CenterNet基础上添加了一个预测角度的分支作为baseline, 通过hm, angle, size和offset 得到obb 的结果。添加angle 分支与size分支几乎一样，loss 为L1loss, 权值为0.1。训练的数据集请参考文章描述。

the number of train epoch

@sunpeng96 Thanks for your interest in our work. The training time(epoch) before and after adding the FSM is the same. The baseline method is also trained for 140 epochs. We...

What is the specific structure of Gc

Just one 3x3 Conv layer. The output channel is determined by the size of the kernel you want. To reduce the amount of parameters, we share the kernel weight among...

Why is NMS instead of max pooling for de-redundancy?

Actually, we use both of them.

代码放的半拉可急的，就别再论文里说自己开源，科学院的人也这么沽名钓誉。

十分感谢关注我们的工作。由于我的偷懒没有把所有代码整理好，十分抱歉。目前就把最核心的开源了，可以当做插件来用。后续会继续做好其他的开源工作。

Where are the feature selection module (FSM) and dynamic reﬁnement heads (DRHs) in code ?

Please refer to [here](https://github.com/Anymake/DRN_CVPR2020/blob/6d5c75895d75601ec92cece3cda516c43abe032a/rotation%20conv%20layer/rotation_conv_utils.py#L189) for FSM. The structure of DRHs is simple, we do not release the code. Please refer to [ issue #20](https://github.com/Anymake/DRN_CVPR2020/issues/20)

When can you publish the code?

Thanks for your attention. W have recently made new progress on rotation convolution and all the energy is spent on it. To make it convenient, we plan to release the...