Clarification on Coordinate Scaling in "get_flow_now_k"

[syt06007]

Hi, thank you for sharing this great repository.

While studying the code, I came across a line in the get_flow_now_k function that left me a bit confused:

four_point = four_point / k

From what I understand, four_point_disp is predicted by the update_block_4 module, which takes as input the correlation and flow—both of which are computed at the feature map resolution. Based on that, I assumed that four_point_disp is also defined in the coordinate space of the feature map.

However, in the code above, four_point is divided by k, which made me wonder if it's actually being treated as if it's in the original image resolution instead.

If my understanding is correct, then the division by k might not be necessary. But I may be missing something, so I was hoping you could clarify the intended coordinate system and whether the division is required.

Thanks in advance for your help!

[syt06007]

Just to follow up—could the reason for dividing four_point by k be that the supervision signal is defined at the original image resolution, rather than at the feature map resolution? If that's the case, it would make sense to scale the predicted displacement accordingly before applying the transformation.