论文阅读 【CVPR-2022】 Accelerating DETR Convergence via Semantic-Aligned Matching

论文阅读 【CVPR-2022】 Accelerating DETR Convergence via Semantic-Aligned Matching

通过语义一致的匹配加快DETR的收敛速度

studyai.com

搜索论文: Accelerating DETR Convergence via Semantic-Aligned Matching

摘要(Abstract)

The recently developed DEtection TRansformer (DETR) establishes a new object detection paradigm by eliminating a series of hand-crafted components.

最近开发出的DEtection TRansformer（DETR）通过消除一系列手工制作的组件建立了一个新的物体检测范式。

However, DETR suffers from extremely slow convergence, which increases the training cost significantly.

然而，DETR存在收敛速度极慢的问题，这大大增加了训练成本。

We observe that the slow convergence is largely attributed to the complication in matching object queries with target features in different feature embedding spaces.

我们观察到，收敛速度慢主要是由于在不同的特征嵌入空间中，物体查询与目标特征的匹配比较复杂。

This paper presents SAM-DETR, a Semantic-Aligned-Matching DETR that greatly accelerates DETR’s convergence without sacrificing its accuracy.

本文提出了SAM-DETR，一种语义对齐匹配的DETR，它大大加快了DETR的收敛速度而不牺牲其准确性。

SAM-DETR addresses the convergence issue from two perspectives.

SAM-DETR从两个角度解决收敛问题。

First, it projects object queries into the same embedding space as encoded image features, where the matching can be accomplished efficiently with aligned semantics.

首先，它将物体查询投射到与编码图像特征相同的嵌入空间，在那里可以通过对齐语义有效地完成匹配。

Second, it explicitly searches salient points with the most discriminative features for semantic-aligned matching, which further speeds up the convergence and boosts detection accuracy as well.

其次，它显式的地用最有鉴别力的特征搜索显著点，进行语义对齐的匹配，这进一步加快了收敛速度，也提高了检测精度。

Being like a plug and play, SAMDETR complements existing convergence solutions well yet only introduces slight computational overhead.

作为一种即插即用的方法，SAMDETR很好地补充了现有的收敛解决方案，但只引入了轻微的计算开销。

Extensive experiments show that the proposed SAM-DETR achieves superior convergence as well as competitive detection accuracy. The implementation codes are publicly available at https://github.com/ZhangGongjie/SAM-DETR .

大量的实验表明，所提出的SAM-DETR实现了卓越的收敛性和有竞争力的检测精度。实施代码可在https://github.com/ZhangGongjie/SAM-DETR 上公开获取。