Question about IBE module

[jimmy133719]

Hi, thanks for your work.

I'm having difficulty understanding how local descriptors integrate gradient cues around shared depth-wise convolutions.
In Formula 3, it appears that the weights of the depth-wise convolution are summed rather than its gradients. I'm curious whether this summation of weights is related to capturing gradient cues.