基于深度学习的快速二进制局部特征
首发时间:2020-03-30
摘要:局部特征是许多视觉任务的基础,随着深度学习的快速发展,大量基于深度学习的局部特征算法被提出。这些算法虽然提升了局部特征的鲁棒性,但是运行速度却十分缓慢。针对此问题,本文提出了一种基于深度学习的快速二进制局部特征。一方面,通过无池化的网络设计,降低了网络深度,实现了特征提取阶段的加速,另一方面,通过特殊的隐层设计,将特征量化为二进制表示,实现了特征匹配阶段的加速。本文使用MegaDepth数据集构建像素级监督对模型进行训练,并在牛津建筑数据集上进行了测试。结果表明,本文提出的局部特征,在保持与其他深度局部特征精度相当的同时,速度方面具有明显优势。
For information in English, please click here
Fast binary local features based on deep learning
Abstract:Local feature is the core of many visual tasks. With the rapid development of deep learning, a large number of local feature algorithms based on deep learning have been proposed. Although these algorithms improve the robustness of local features, they run very slowly. To solve this problem, this paper proposes a fast binary local feature based on deep learning. On the one hand, the network depth is reduced by the no-pooling design, and the feature extraction stage is accelerated; on the other hand, the feature is quantified into binary representation by the special hidden layer design, and the feature matching stage is accelerated. This paper uses the MegaDepth dataset to build pixel-level supervision to train the model and tests it on the Oxford building dataset. The results show that the local feature presented in this paper has obvious advantages in speed while maintaining the same accuracy as other deep local features.?
Keywords: local feature deep learning fast binary
基金:
引用
No.****
同行评议
勘误表
基于深度学习的快速二进制局部特征
评论
全部评论0/1000