How do I know which batch of feature is input to the bbox_head?

[babajakpum]

Hello, all!

I am new to this area and mmdetection, so please kindly understand if I don't get your advice.

I have a question about mask rcnn.
Output of the backbone and neck is x (tuple[Tensor]).
The tuple length is 5, and the shape of the tensor is torch.Size([2, 256, 336, 192]).
So far, I can see that it includes batch information, which is 2.

In the following code, x is then input to self.bbox_roi_extractor along with rois.
rois is a tensor with the shape (batch n, 5). However, the shape of bbox_feats, which is input to the bbox_head, is torch.Size([1024, 256, 7, 7]). I don't see any batch information in bbox_feats. Furthermore, in bbox_head, the feature becomes a 1024 * 1024 tensor after some processing.

Why doesn't it contain any batch information? Could you please help me understand?

bbox_feats = self.bbox_roi_extractor(x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats, y)