Possible Normalization Bug in Deformable Attention Coordinates

[Ali2500]

This pertains to be cross-attention between queries and encoder feature maps in deformable_transformer.py. The reference_points for the deformable attention operation appear to be interpreted as (x,y) coordinates since they are multiplied by src_valid_ratios in line 339 which has a (width, height) format (as evident from the get_valid_ratio method, line 125).

However, the src_spatial_shapes tensor contains dimensions in (height, width) format (see line 138 and 139). Then, in ms_deform_attn.py the reference points and offsets are combined as follows:

sampling_locations = reference_points[:, :, None, :, None, :] \
                       + sampling_offsets / input_spatial_shapes[None, None, None, :, None, :]

So it appears that the sampling_offsets / input_spatial_shapes[None, None, None, :, None, :] part assumes coordinates to be in (y,x) format whereas reference_points[:, :, None, :, None, :] appears to be in (x,y) format.

Of course I'm not sure if I followed the code correctly here (I could have missed something), but just wanted to bring this to the authors' attention in case they did not know already.

[jackroos]

@Ali2500 Thank you for your kind reminder. This is indeed a mistake made in this released version, but not in the version used in our paper. I will update the code and pre-trained models soon. Thanks again. I will close this issue after the code is updated.