problem about outputs_coord

[luotianhang]

outputs_coord_delta = self.bbox_embed(hs)
outputs_coord_cxcy = outputs_coord_delta[..., :2] * ref_unsigmoid[..., 2:] + ref_unsigmoid[..., :2]
outputs_coord_wh = outputs_coord_delta[..., 2:].exp() * ref_unsigmoid[..., 2:]
outputs_coord = torch.concat(
[outputs_coord_cxcy, outputs_coord_wh], dim=-1
)

I am working on this project ,and find a question ,the ref_unsigmoid is very unstable, and how to make sure all the value of outputs_coord is in [0,1],without a sigmoid

[isaacrob]

This is a fairly common reparameterization of bounding box regression in modern DETRs. It doesn't guarantee the coordinates are in 0-1, but has cleaner gradients. You can always clamp them in post processing.