专栏首页arxiv.org翻译专栏用于移动设备高性能图像匹配任务的卷积神经网络压缩(CS CV)
原创

用于移动设备高性能图像匹配任务的卷积神经网络压缩(CS CV)

深度神经网络已经证明了最先进的性能,基于特征的图像匹配,通过新的大型和多样化的数据集的出现。然而,在评估这些模型的计算成本、模型大小和匹配精度权衡方面几乎没有工作。本文通过考虑最先进的L2Net体系结构,明确地解决了这些实际约束问题。我们观察到L2Net体系结构中存在显著的冗余,我们通过使用深度可分层和高效的Tucker分解来开发这种冗余。我们证明,这些方法的组合更有效,但仍然牺牲了高端的准确性。因此,我们提出了卷积-Dept wise-Point wise(CDP)层,它提供了一种在标准卷积和深度可分卷积之间进行插值的方法。有了这个提议的层,我们能够在L2Net体系结构上实现多达8倍的参数减少,13倍的计算复杂度减少,同时牺牲不到1%的整体精度跨越HPATCH基准。为了进一步演示这种方法的推广,我们将其应用于Super Point模型。我们表明,CD P层提高了精度,同时使用明显较少的参数和浮点运算进行推理。

原文:Deep neural networks have demonstrated state-of-the-art performance for feature-based image matching through the advent of new large and diverse datasets. However, there has been little work on evaluating the computational cost, model size, and matching accuracy tradeoffs for these models. This paper explicitly addresses these practical constraints by considering state-of-the-art L2Net architecture. We observe a significant redundancy in the L2Net architecture, which we exploit through the use of depthwise separable layers and an efficient Tucker decomposition. We demonstrate that a combination of these methods is more effective, but still sacrifices the top-end accuracy. We therefore propose the Convolution-Depthwise-Pointwise (CDP) layer, which provides a means of interpolating between the standard and depthwise separable convolutions. With this proposed layer, we are able to achieve up to 8 times reduction in the number of parameters on the L2Net architecture, 13 times reduction in the computational complexity, while sacrificing less than 1% on the overall accuracy across the HPatches benchmarks. To further demonstrate the generalisation of this approach, we apply it to the SuperPoint model. We show that CDP layers improve upon the accuracy while using significantly less parameters and floating-point operations for inference.

原文题目:Compression of convolutional neural networks for high performance imagematching tasks on mobile devices

原文作者:Roy Miles, Krystian Mikolajczyk

原文地址:https://arxiv.org/abs/2001.03102

原创声明,本文系作者授权云+社区发表,未经许可,不得转载。

如有侵权,请联系 yunjia_community@tencent.com 删除。

我来说两句

0 条评论
登录 后参与评论

相关文章

  • 循环热管的动态状态空间建模与基于模型的控制设计(CS SY)

    对于航空航天、汽车或服务器系统中电子元件的热控制,散热器通常远离热源。因此,热传导系统是有效冷却电子元件所必需的。循环热管(LHPs)就是这样的传热系统,它利用...

    用户6853689
  • 芬兰语建模与深层变压器模型(CS SD)

    在LSTM被认为是主导模型体系结构之后的很长一段时间,转换器在语言建模中占据了中心舞台。在这个课题中,我们研究了BRET转换器结构和XL转换器结构在语言建模任务...

    用户6853689
  • 音频修复:重新访问和重新加权(CS SD)

    我们处理了稀疏的音频修复问题。优化方法之一是在填充的间隙内信号的能量不足。我们提出了基于稀疏性和凸优化的音频嵌入框架的改进方案,以补偿这种能量损失。新的思想是基...

    用户6853689
  • Common Pitfalls to Avoid when using HTML5 Application Cache

    Application Cache, also known as AppCache, has been a pretty hot topic with web ...

    IMWeb前端团队
  • 多目标进化算法应用于提高医药数据领域学习器的性能(CS AI)

    原文标题完整翻译:多目标进化算法应用于提高在医药数据领域使用整体特征选择和离散化模型的学习器的性能

    Donuts_choco
  • poj-------------(2752)Seek the Name, Seek the Fame(kmp)

    Seek the Name, Seek the Fame Time Limit: 2000MS Memory Limit: 65536K Tot...

    Gxjun
  • POJ-1926 Pollution

    Pollution Time Limit: 1000MS Memory Limit: 30000K Total Submissions: 40...

    ShenduCC
  • Common Pitfalls to Avoid when using HTML5 Application Cache

    本文作者:IMWeb 黎腾 原文出处:IMWeb社区 未经同意,禁止转载 Application Cache, also known as AppCa...

    IMWeb前端团队
  • POJ-2029 Get Many Persimmon Trees(动态规划)

    Get Many Persimmon Trees Time Limit: 1000MS Memory Limit: 30000K Total ...

    ShenduCC
  • Common Pitfalls to Avoid when using HTML5 Application Cache

    Application Cache, also known as AppCache, has been a pretty hot topic with web ...

    IMWeb前端团队

扫码关注云+社区

领取腾讯云代金券