Scene graph descriptors for visual place classification from noisy scene data
Scene graph descriptors for visual place classification from noisy scene data
复制标题
用于从噪声场景数据中进行视觉位置分类的场景图描述符
DOI：
10.1016/j.icte.2022.11.003
复制
发表时间：
2023
期刊：
影响因子：
5.4
通讯作者：
Ryo Yamamoto
中科院分区：
文献类型：
--
作者：
Tomoya Ohta;Kanji Tanaka;Ryo Yamamoto
在视觉机器人位置识别（VPR）中，场景图是一种丰富的场景模型，它可以描述场景中复杂的上下文，如各种视觉内容之间的关系，包括外观、空间和语义。然而，训练一个有效的场景图分类器并不简单。现有的方法通常依赖于查询和数据库图之间的穷举匹配，并且不能扩展到大规模的VPR问题。我们的研究的动机是最近发展的图卷积神经网络（GCN）作为一个有效的和有区别的分类器的图形数据，它的目的是探索的GCN作为一个场景图分类器的潜力。然而，不像几个现有的GCN应用程序，没有有效的场景图描述符的GCN分类噪声场景数据存在。为了解决这个问题，在这里，我们建议通过采用现有的最先进的单视图VPR系统作为教师模型，在教师到学生的知识转移计划中训练GCN模型。所提出的方法是在一个实际的VPR框架内实现的，通过结合以下三个独立的领域：多模态信息检索，排名匹配，和基于相似性的模式识别。在公开的NCLT数据集上的实验验证了该方法的有效性。
In visual robot place recognition (VPR), a scene graph is a rich scene model that can describe the complex contexts in a scene such as the relationships between various types of visual contents including appearance, space, and semantics. However, training an efficient scene graph classifier is not straightforward. Existing approaches typically rely on exhaustive matching between query and database graphs and are not scalable to large-size VPR problems. Our research is motivated by a recent development of the graph convolutional neural network (GCN) as an efficient and discriminative classifier for graph data, and it aims to explore the potential of the GCN as a scene graph classifier. However, unlike several existing GCN applications, no valid scene graph descriptor for a GCN classifier on noisy scene data exists. To address this issue, herein, we propose to train the GCN model in a teacher-to-student knowledge transfer scheme by employing an existing state-of-the-art single-view VPR system as the teacher model. The proposed approach is implemented within a practical VPR framework by combining the best of the following three independent fields: multimodal information retrieval, rank matching, and similarity-based pattern recognition. Experiments using the public NCLT dataset validate the effectiveness of the proposed approach.