当前位置: X-MOL 学术Pattern Anal. Applic. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
SSD based on contour–material level for domain adaptation
Pattern Analysis and Applications ( IF 3.7 ) Pub Date : 2021-05-11 , DOI: 10.1007/s10044-021-00986-w
Ning Jiang , Jinglong Fang , Jihui Xu , Yanli Shao

In the field of object detection, domain migration has gradually become a hot issue. We hope that the model trained in the domain with label-rich can be applied to other domains with label-poor or without labels, which can save a lot of time and energy for annotation, but different domain distributions are always mismatched; such a distribution mismatch will lead to a sharp decline in domain transfer performance. In this work, to improve the performance of object detection for domain transfer, we tackle the domain shift on two levels: (1) the contour-level shift, such as appearance, shape, and size, and (2) the material-level shift, such as texture, shade, and color. We apply different alignments to the aforementioned levels, specifically contour-level adaptation with full alignment and material-level adaptation with selective alignment. We construct a domain adaptation framework based on the recent state-of-the-art SSD model, and SSD is the abbreviation of single shot multibox detector, which is preeminent above most of the other approaches proposed in large numbers due to its real-time performance and effectiveness. We design two domain adapters on contour level and material level, respectively, to alleviate the domain discrepancy. Recently, approaches that align distributions of source and target images employing an adversarial loss have been proven effective, so the two domain adapters are implemented by learning a domain classifier in adversarial training manner, and the domain classifiers on different levels are further reinforced with a consistency regularization in the SSD model. We empirically verify the effectiveness of our method, which outperforms the other three state-of-the-art methods by a large margin of 5–10% in terms of mean average precision (mAP) on various datasets in both similar and dissimilar domain shift scenarios.



中文翻译:

基于轮廓-材料级别的SSD,用于领域适应

在对象检测领域,域迁移已逐渐成为一个热门问题。我们希望在标签丰富的领域中训练的模型可以应用于标签贫乏或没有标签的其他领域,这样可以节省大量的时间和精力进行标注,但是不同的领域分布总是不匹配的;这样的分配不匹配将导致域转移性能急剧下降。在这项工作中,为了提高用于区域转移的对象检测的性能,我们在两个级别上解决了区域偏移:(1)轮廓级别的偏移,例如外观,形状和大小,以及(2)材质级别移位,例如纹理,阴影和颜色。我们对上述级别应用了不同的对齐方式,特别是轮廓级别的完全对齐和材质级别的选择性对齐。我们基于最新的最新SSD模型构建域适应框架,SSD是Single Shot Multibox Detector的缩写,由于它具有实时性,因此它比其他大多数其他方法都优越性能和有效性。我们分别在轮廓级别和材质级别设计了两个域适配器,以缓解域差异。最近,已经证明利用对抗损失来对准源图像和目标图像的分布的方法是有效的,因此通过以对抗训练的方式学习域分类器来实现两个域适配器,并且在一致性上进一步增强了不同级别的域分类器。 SSD模型中的正则化。我们凭经验验证了我们方法的有效性,

更新日期:2021-05-11
down
wechat
bug