人工智能培训

搜索

人工智能论文:面向动态识别的双流视频超分辨率(Two-Stream Oriented Video Super-Resolution for Action Rec

[复制链接]
bigrc 发表于 2019-3-15 12:59:24 | 显示全部楼层 |阅读模式
bigrc 2019-3-15 12:59:24 223 0 显示全部楼层
人工智能论文:面向动态识别的双流视频超分辨率(Two-Stream Oriented Video Super-Resolution for Action Recognition)我们研究视频超分辨率(SR)问题不是为了视觉质量,而是为了促进视频分析任务,例如行动认可。基于卷积网络的人口识别方法,以双流网络为例,不能直接应用于不同空间分辨率的视频。这可以通过在识别之前执行视频SR来解决,这促使我们改进用于识别准确性的SR过程。针对双流动作识别网络,我们分别针对空间和时间流提出两个视频SR方法。一方面,我们认为图像SR方法的附加细节可能对识别有帮助或有害,我们为空间导向SR(SoSR)网络提出了光流导向加权平方误差损失。另一方面,我们观察到现有的视频SR方法在帧之间产生临时连续性,这也使识别准确性恶化,并且我们为时间导向SR(ToSR)提出了一个暹罗网络,强调了连续帧之间的时间连续性。我们使用两个最先进的动作识别网络和两个已知的数据集(UCF101和HMDB51)进行实验。结果证明了我们提出的SoSR和ToSR在提高识别准确性方面的有效性。
We study the video super-resolution (SR) problem not for visual quality, butfor facilitating video analytics tasks, e.g.action recognition.The popularaction recognition methods based on convolutional networks, exemplified bytwo-stream networks, are not directly applicable on videos of different spatialresolutions.This can be remedied by performing video SR prior to recognition,which motivates us to improve the SR procedure for recognition accuracy.Tailored for two-stream action recognition networks, we propose two video SRmethods for the spatial and temporal streams respectively.On the one hand, weobserve that the added details by image SR methods can be either helpful orharmful for recognition, and we propose an optical-flow guided weightedmean-squared-error loss for our spatial-oriented SR (SoSR) network.On theother hand, we observe that existing video SR methods incur temporaldiscontinuity between frames, which also worsens the recognition accuracy, andwe propose a siamese network for our temporal-oriented SR (ToSR) thatemphasizes the temporal continuity between consecutive frames.We performexperiments using two state-of-the-art action recognition networks and twowell-known datasets--UCF101 and HMDB51.Results demonstrate the effectivenessof our proposed SoSR and ToSR in improving recognition accuracy.人工智能论文:面向动态识别的双流视频超分辨率(Two-Stream Oriented Video Super-Resolution for Action Recognition) qIrzzNNGwna2Xagn.jpg
URL地址:https://arxiv.org/abs/1903.05577     ----pdf下载地址:https://arxiv.org/pdf/1903.05577    ----人工智能论文:面向动态识别的双流视频超分辨率(Two-Stream Oriented Video Super-Resolution for Action Recognition)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

bigrc当前离线
新手上路

查看:223 | 回复:0

快速回复 返回顶部 返回列表