how to understand predicting mask_embed using only query_feat

[gaosanyuan]

In the prediction head of Mask2Former, the mask embedding is solely computed based on the decoder output, specifically the query features (query_feat). This raises a question about the model's ability to differentiate between objects with similar semantics but different spatial locations, as they might share similar query features. It's unclear why the model doesn't incorporate additional information, such as dedicated query embeddings, to enhance the discrimination process.
Thanks