About bbox refinement

[satuoqaq]

bbox_pred = self.deform_sampling(decoded_bbox_preds.contiguous(), reg_offset.contiguous())

bbox_pred = F.relu(bbox2distance(points, bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4)).reshape(b, h, w, 4).permute(0, 3, 1, 2).contiguous())

first line you sample in coarse bbox pred , i think this is final pred, why you decode again in second line?

[Icecream-blue-sky]

bbox_pred = self.deform_sampling(decoded_bbox_preds.contiguous(), reg_offset.contiguous())

will return the absolute coordinate of bounding boxes. The second line transforms it to the [l,t,b,r] style.

[satuoqaq]

bbox_pred = self.deform_sampling(decoded_bbox_preds.contiguous(), reg_offset.contiguous())

will return the absolute coordinate of bounding boxes. The second line transforms it to the [l,t,b,r] style.

my fault, I thought it was distance to bbox again, i got it~