MDS-DETR: DETR with Masked Duplicate Suppressor

Researchers introduced MDS-DETR, a DETR-style object detector that combines one-to-one and one-to-many supervision in a single decoder.

It uses confidence-based causal masking to suppress duplicate predictions during decoding, improving recall and precision.

On MS COCO, MDS-DETR outperformed several DETR variants, including Deformable-DETR, MS-DETR, and Relation-DETR, with only modest training overhead.

The approach improves convergence and accuracy without adding auxiliary decoders, making end-to-end deployment cleaner and more efficient.