Deformable Convolution Networks introduce learnable offsets to improve convolution operations, enhancing image classification and detection tasks. The networks adapt to object shapes, providing better results compared to traditional fixed convolution methods. By predicting offsets specific to each spatial point, deformable convolutions can effectively capture fine-grained action and object details, reducing background noise."
133051545,19974094,Python字典转换技巧：键值互换与内容修饰,"['Python', '数据结构', '字典操作']
文章列表：
Active Convolution: Learning the Shape of Convolution for Image Classification
这篇文章首先将卷积进行参数的拓展，之前都是直接学习权重就是了。之前的卷积都是fixed的采样。比如3*3，就是采样与window中心点相对位置一样的9个点。即：
Y=W∗X+bY=W*X+bY=W∗X+b
ym,n=∑c∑i,jwc,i,j⋅xc,m+i,n+j+by_{m,n}=\sum_{c}\sum_{i,j}w_{c,i,j}\cdot x_{c,m+i,n+j}+bym,n=c∑i,j∑wc,i,j⋅xc,m+i,n+j+b
如果将XXX进行修改：
Y=W∗Xθp+bY=W*X_{\theta_{p}}+bY=W∗Xθp+b
ym,n=∑c∑kwc,k⋅xc,m+αk,n+βk+by_{m,n}=\sum_{c}\sum_{k}w_{c,k}\cdot x_{c,m+ \alpha _{k}, n+\beta_{k}}+by
最低0.47元/天 解锁文章
扫一扫
8325
到【灌水乐园】发言
