PF-WILLOW

在存在类内变化和场景布局发生较大变化的情况下，寻找图像对应仍然是一个具有挑战性的问题。语义流方法旨在处理描绘同一对象或场景类别的不同实例的图像。我们引入了一种新的语义流方法，称为提案流，它使用对象提案建立可靠的对应关系。与在像素或定期采样的局部区域上运行的流行语义流方法不同，提案流受益于现代对象提案的特征，在多个尺度上表现出高可重复性，并且可以利用提案之间的局部和几何一致性约束。我们还表明，相应的稀疏提议流可以有效地转换为传统的密集流场。我们引入了两个新的具有挑战性的数据集，可用于评估通用语义流技术和基于区域的方法，例如提案流。我们使用这些基准将提案流中的不同匹配算法、对象提案和区域特征与语义流中的最新技术进行比较。这种比较以及对标准数据集的实验表明，提案流在各种设置中明显优于现有的语义流方法。

Finding image correspondences remains a challenging problem when there exist intra-class variations and significant changes in scene layout. Semantic flow methods aim to handle images that depict distinct instances of the same object or scene category. We introduce a novel semantic flow method termed ProposalFlow, which establishes reliable correspondences using object proposals. Unlike prevailing semantic flow methods that operate on pixels or regularly sampled local regions, ProposalFlow leverages the strengths of modern object proposals, which exhibit high repeatability across multiple scales, and can exploit local and geometric consistency constraints between proposals. We further demonstrate that the corresponding sparse proposal flow can be efficiently transformed into conventional dense flow fields. We introduce two new challenging datasets that can be used to evaluate general semantic flow techniques and region-based methods such as ProposalFlow. We use these benchmarks to compare different matching algorithms, object proposals, and region features within ProposalFlow against state-of-the-art semantic flow methods. This comparison, alongside experiments on standard datasets, demonstrates that ProposalFlow outperforms existing semantic flow methods across various settings.

搜集汇总

背景与挑战

背景概述

PF-WILLOW是一个专注于图像语义对应问题的数据集，特别设计用于评估在类内和场景布局变化大的情况下的语义流技术。该数据集支持提案流方法的研究，该方法通过对象提案建立图像对应关系，并在实验中表现优于现有方法。