Jiaming Han

AlignConv依赖deformable conv， ncnn未实现DCN，不能直接部署

没有特殊处理，仅是沿用[OWOD](https://github.com/JosephKJ/OWOD)的设定

请链接具体的代码位置？

Hi~ Did you adjust the learning rate and total training iterations, or you just train the model with `batch size=16` with 1 GPU?

Can you reproduce the results of baseline method `Faster R-CNN`?

Sure. We use N_cls=82 just for the compatibility of ORE.

It is the output of bbox head (RoI Align+FCs). We add a hook to [roi_heads.py](https://github.com/csuhan/opendet2/blob/main/opendet2/modeling/roi_heads/roi_heads.py). ```python def _forward_box(self, features: Dict[str, torch.Tensor], proposals: List[Instances], targets=None): ##### # remove some lines #####...

`FR-CNN` is a standard FR-CNN, while `baseline` denotes a variant of FR-CNN with some modifications (see Sec. 3.2). Although the `baseline` does not improve the open-set performance, it works better...

1. 我们的训练是在20-class的VOC上进行的，测试是在80-class的VOC+COCO上进行的。因此训练时尽可能地避免了open-set样本的出现（当然还是会存在一些open-set样本，即VOC中未标注的non-VOC class） 2. 理论上没有影响，即这个设定与open-set无关。但close-set的所有类当作一个类别训练可能会忽略类别之间的语义信息，进而影响close-set与open-set类之间的关系。

Sorry. Most code of `opendet2/solver/build.py` is copied from Detectron2 (including line 39). Can you check the version of your detectron2? We tested `v0.5` and it works well.