中文MDS-DETR：带掩码重复抑制器的DETR

MDS-DETR提出联合使用一对一和一对多匹配监督，解决DETR收敛慢、召回率低的问题。传统方法虽增加一对多辅助目标但训练成本高且推理时丢弃辅助解码器。本方法在训练中保留所有解码器，有效提升性能且不增加推理开销，实现更高效的目标检测。

arXiv:2605.23507v1 Announce Type: new Abstract: The DEtection TRansformer (DETR) is a powerful end-to-end object detector, yet its one-to-one matching strategy suffers from slow convergence and low recall. A common approach to address this issue is to use one-to-many label assignment to provide more positive samples. However, existing methods that use one-to-many matching as an auxiliary objective lead to increased training costs, with their auxiliary decoders discarded during inference. To address this limitation, we propose MDS-DETR, which leverages both one-to-one and one-to-many supervisio