论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks

Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks for Fine-Grained Action Detection

Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019

task

细粒度的动作检测

阅读记录

论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第1张图片
文中认为,如果用deformable卷积,那么两帧之间的offsets的差值就是光流。因为,如果deformable conv用于动作识别中,这个deformable如果学习得比较好,那么相当于一种检测关键点的功能。那么显然,这些关键点的变化,就是光流。
论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第2张图片
(a)和(b)是两个连续的帧;
©和(d)为背景和运动区域的运动矢量(绿点为激活位置,红箭头为运动矢量);
(e)为该人在t - 1和t时间手动定义的mask
(f)为特征空间中运动场的能量,通过聚合所有可变形卷积层中的运动向量计算得到。
论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第3张图片
论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第4张图片
论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第5张图片论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第6张图片论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第7张图片论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第8张图片论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第9张图片论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第10张图片论文阅读 Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks_第11张图片

说明

以上内容均为作者本人平时阅读并且汇报使用,内容整理全凭个人理解,如有侵权,请联系我;内容如有错误,欢迎留言交流。转载请注明出处,并附有原文链接,谢谢!

此外,我还喜欢用ipad对论文写写画画(个人英文阅读的水平有限),做一些断句、重点勾画等,有兴趣大家可以按需下载:链接

更多论文分享,请参考: 深度学习相关阅读论文汇总(持续更新)

你可能感兴趣的:(深度学习论文阅读,深度学习,python,神经网络,计算机视觉,cnn)