Unnecessary slots tend to over-segment?

[franciscocms]

Hey! I guess this could be a dumb question, but I honestly couldn't answer this by analyzing your paper. Do you have a clear idea about what happens to slots that are not required to explain the input image? For instance, in the original SA algorithm, it is clear that unnecessary slots tend to encode background information, but this does not happen with SA-MESH. It seems like some slots end up "binding" to regions that belong to already encoded objects (e.g. in Figure 7, in images with objects that are not strongly occluded). Since attention masks are sparse, it seems like this over-segmentation issue is easier to occur. Am I wrong? Thanks in advance!