改写这段文字This paper investigates a phenomenon where query- based object detectors mispredict at the last decoding stage while predicting correctly at an intermediate stage. We review the training process and attribute the overlooked phenomenon to two limitations: lack of training emphasis and cascading errors from decoding sequence. We design and present Selective Query Recollection (SQR), a simple and effective training strategy for query-based object detec- tors. It cumulatively collects intermediate queries as decod- ing stages go deeper and selectively forwards the queries to the downstream stages aside from the sequential struc- ture. Such-wise, SQR places training emphasis on later stages and allows later stages to work with intermediate queries from earlier stages directly. SQR can be easily plugged into various query-based object detectors and sig- nificantly enhances their performance while leaving the in- ference pipeline unchanged. As a result, we apply SQR on Adamixer, DAB-DETR, and Deformable-DETR across various settings (backbone, number of queries, schedule) and consistently brings 1.4 ∼ 2.8 AP improvement.

本文研究了一种现象，即在预测中间阶段正确的情况下，在最后的解码阶段出现查询基础目标检测器误判的情况。我们回顾了训练过程，并将被忽视的现象归因于两个限制：缺乏训练重点和解码序列中的级联错误。我们设计并提出了选择性查询回收（SQR），这是一种简单而有效的查询基础目标检测器训练策略。随着解码阶段的加深，它累积收集中间查询，并选择性地将查询转发到下游阶段，除了顺序结构之外。因此，SQR将训练重点放在后期阶段，并允许后期阶段直接使用早期阶段的中间查询。SQR可以轻松地插入各种查询基础目标检测器，并显着提高它们的性能，同时保持推理管道不变。因此，我们在各种设置（骨干网络，查询数量，时间表）上将SQR应用于Adamixer，DAB-DETR和Deformable-DETR，并始终带来1.4〜2.8 AP的改进。

相关问题

试着改述这一段英语This paper investigates a phenomenon where query- based object detectors mispredict at the last decoding stage while predicting correctly at an intermediate stage. We review the training process and attribute the overlooked phenomenon to two limitations: lack of training emphasis and cascading errors from decoding sequence. We design and present Selective Query Recollection (SQR), a simple and effective training strategy for query-based object detec- tors. It cumulatively collects intermediate queries as decod- ing stages go deeper and selectively forwards the queries to the downstream stages aside from the sequential struc- ture. Such-wise, SQR places training emphasis on later stages and allows later stages to work with intermediate queries from earlier stages directly. SQR can be easily plugged into various query-based object detectors and sig- nificantly enhances their performance while leaving the in- ference pipeline unchanged. As a result, we apply SQR on Adamixer, DAB-DETR, and Deformable-DETR across various settings (backbone, number of queries, schedule) and consistently brings 1.4 ∼ 2.8 AP improvement.

这篇论文研究了一个现象，即在预测中间阶段正确的情况下，基于查询的物体检测器在最后的解码阶段误判。我们回顾了训练过程，并将这种被忽视的现象归因于两个限制：缺乏训练重点和解码序列中的级联错误。我们设计并提出了选择性查询回收（SQR），这是一种简单而有效的基于查询的物体检测器训练策略。随着解码阶段的深入，SQR累积收集中间查询，并选择性地将查询转发到下游阶段，除了顺序结构之外。这样，SQR将训练重点放在后期阶段，并允许后期阶段直接使用来自早期阶段的中间查询。SQR可以轻松地插入各种基于查询的物体检测器，并显著提高它们的性能，同时不改变推理流程。因此，我们在各种设置（骨干网络，查询数量，时间表）上将SQR应用于Adamixer，DAB-DETR和Deformable-DETR，并始终带来1.4〜2.8 AP的改进。

Abstract—In heterogeneous networks (HetNets), user association approaches should be able to achieve load balancing among base stations (BSs). This paper investigates the joint optimization of user association and resource allocation in Backhaul-constrained HetNets for capacity enhancements. We consider two major limitations in HetNets: the backhaul bottleneck of BSs and the capability of user equipment (UE). We establish a framework based on a multi-leader multi-follower Stackelberg game, in which resource allocation is formulated as a follower-level game and user association is cast as a leader-level game. Because of the backhaul bottleneck of small BSs, the given preference order of users renders the final association result unstable. Thus, the resident-oriented GaleShapley (GS) algorithm is included in the proposed framework to obtain a stable single-BS association. Furthermore, congestion factors are introduced to reflect the relative backhaul congestion degrees of BSs, which enables load balancing among the small BSs in the proposed algorithm. The study considers user association and resource allocation with and without limitations on the number of serving users for small BSs in HetNets. Extensive simulation results suggest that the proposed algorithm can adaptively respond to a wide variety of network situations.中文

在异构网络（HetNets）中，用户关联方法应该能够实现基站（BS）之间的负载平衡。本文研究了在受限于回程链路的HetNets中，用户关联和资源分配的联合优化问题，以增强系统容量。我们考虑了HetNets中的两个主要限制：BS的回程瓶颈和用户设备（UE）的能力。我们建立了一个基于多领导者-多追随者Stackelberg博弈的框架，其中资源分配被形式化为追随者级别的博弈，用户关联被视为领导者级别的博弈。由于小型BS的回程瓶颈，给定的用户优先顺序导致最终关联结果不稳定。因此，该提出的框架中包括了面向居民的Gale-Shapley（GS）算法，以获得稳定的单BS关联。此外，引入了拥塞因子来反映BS的相对回程拥塞程度，从而实现了提出算法中小型BS之间的负载平衡。本研究考虑了HetNets中限制小型BS服务用户数量和不限制的用户关联和资源分配。广泛的仿真结果表明，该提出的算法能够适应各种网络情况。