If I understood the paper correctly, SwAV encourages the cluster assignments between the different views from the 'same' image to be similar.
Then, is there any structure on the assignments between 'different' images?
For example, are similar images (e.g. images from the same class) assigned to similar clusters?
If so, what part of the algorithm encourages this behavior? Or is it just happening empirically?
Hi, thanks for the great work!
If I understood the paper correctly, SwAV encourages the cluster assignments between the different views from the 'same' image to be similar.
Then, is there any structure on the assignments between 'different' images? For example, are similar images (e.g. images from the same class) assigned to similar clusters? If so, what part of the algorithm encourages this behavior? Or is it just happening empirically?