Not able to understand the offset

[pmgautam]

According to fig 2 in the paper, offset is causing the kernel to look into other positions than the regular 9 points in the case of 3x3 convolution.

However, when bilinear interpolation is used from eq. 3, it is computing the input feature to be used in the fractional position with the kernel instead of the new positions to look into. I'm confused from these differences regarding the mechanism of offset computation.

From the equations I see that we are not computing offset for the positions, instead we are computing the new feature by interpolating over the existing features. But from the diagram it looks to be computing the new offset positions and using the features from those positions. Can someone help me understand this. Any help is highly appreciated.