【DeepLearning】用于几何匹配的卷积神经网络体系结构

【论文标题】Convolutional neural network architecture for geometric matching (2017CVPR)

【论文作者】Ignacio Rocco ,Relja Arandjelovi´,Josef Sivic

【论文链接】Paper (15-pages // Double column)

【Abstract】

We address the problem of determining correspondences between two images in agreement with a geometric model such as an affine or thin-plate spline transformation, and estimating its parameters. The contributions of this work are three-fold. First, we propose a convolutional neural network architecture for geometric matching. The architecture is based on three main components that mimic the standard steps of feature extraction, matching and simultaneous inlier detection and model parameter estimation, while being trainable end-to-end. Second, we demonstrate that the network parameters can be trained from synthetically generated imagery without the need for manual annotation and that our matching layer significantly increases generalization capabilities to never seen before images. Finally, we show that the same model can perform both instance-level and category-level matching giving state-of-the-art results on the challenging Proposal Flow dataset.

我们解决了两个图像之间的对应关系的问题,使用的是一个几何模型,例如仿射或薄板样条变换,并估计其参数。这项工作的贡献有三方面。

首先,我们提出了一个卷积神经网络结构的几何匹配。该架构基于三个主要组件,它们模拟特征提取、匹配和同步的异常检测和模型参数估计的标准步骤,同时可以进行端到端的训练。其次,我们证明了网络参数可以通过综合生成的图像进行训练,且无需人工标注,而且我们的匹配层显著提高了在从未见过图像之前的泛化能力。

最后,我们展示了相同的模型可以同时执行实例级和类别级匹配,为具有挑战性的建议流数据集提供最先进的结果。

上一篇:linux编程中接收主函数返回值以及错误码提示


下一篇:Android TV端的(RecyclerView)水平滚动焦点错乱问题