Question about Cross-MFE and Self-MFE

[hamzarawal]

Hi, thanks for the awesome work. I think are a few differences between the code and the paper. Fig.3 in the paper shows that Self-MFE only takes in the reference features (or warped reference features). And the cross-MFE takes Source features and warped reference features as inputs. However, in the code, Self-MFE takes warped source features and warped reference features (not shown in Fig.3) as inputs as shown below: input_feat = torch.cat([att_source_feat,att_reference_feat],1)

Also, the Cross-MFE takes in the raw reference features (instead of warped as shown in Fig.3) and warped source features of layer n-1. input_feat = torch.cat([att_source_feat,feat_ref],1)

Am I getting this correctly? And is there any particular reason for these changes?

[ShuaiBai623]

1 In our final implementation. Self-MFE takes warped source features and warped reference features as inputs as, and it achieves slightly better performance 2 Since the network regresses the residual flow, warped features are used as inputs .

[hamzarawal]

Thank you. I understand the first answer. For the second point, Cross-MFE is supposed to use warped features as you have said and the diagram shows as well, but in the code un-warped features (feat_ref) are used as you can see in the above line.