Question regarding Relationship Modeling

[yxchng]

According to this paragraph, region-based queries are supervised by mini-map down-sampled from the ground truth. If I understand correctly, all the queries then have the same supervision. If so, how can these queries learn to correspond to different regions? Wouldn't they learn the same thing and correspond to the same region?

Can you kindly explain more about this?

[henghuiding]

Hi,

The region queries have different supervision. Split the image to 10x10 patches, then we have 10x10=100 region queries, and a 10x10 mini-map. If the (3, 6)-th image patch contains object while (6, 8)-th image patch doesn't, the label for the (3, 6)-th region query on the mini-map is set to 1, whereas the label for the (6, 8)-th region query is set to 0. I hope this example clarifies the process for you. Thank you!