ROAM icon indicating copy to clipboard operation
ROAM copied to clipboard

What's the difference between θcf and θ reg?

Open xieyunjiao opened this issue 4 years ago • 0 comments

Hi,thank you for your contribution!I read your article recently. 1.Except the initialization weights θ and the branches of the application are different.What's the difference between correlation filter θcf and bounding box regression filter θreg ? 2.如果θreg和 θcf都用双线性插值重新调整了目标特征图的大小,这跟摘要里提到的the resizable convolutional filters什么关系呀?文章里很多名词都用了resizable 修饰了,他们的主要根源都是因为用了卷积滤波器的双线性插值吗?抱歉我看文章看的有点晕了 3.根据这一段的描述"The tracking model contains two branches where the response generation branch determines the presence of target by predicting a confidence score map and the bounding box regression branch estimates the precise box of the target by regressing coordinate shifts from the box anchors mounted on the sliding-window positions. " 和从文章中的Figure1中看,预测目标存在的 the response generation branch 和the bounding box regression branch是两个平行的分支,这两个分支在元学习框架中的交界点一起更新了损失 image。响应生成分支用来确定帧中目标是否存在的话,这个分支的结果有什么用呀?两个分支是各用各的结果吗,各自更新? 4.当LSTM模型更新的时候,文章中描述到将更新了的θ模型用到未来的随机帧数上测试模型的性能。。。这种随机帧的条件下,如果目标不存在于当前帧中的条件下,即使在未来帧更新了相应的模型θcf and θ reg,此时响应生成分支存在的意义是什么? 5.边框回归分支是从Anchor转换得到精确的矩形框,然后用这些转换来的框做边框回归的输入,进而预测未来帧中的目标边框,文章说由于这个框是resizable的就只设置了一个Anchor,是直接把VGG-16传过来的特征图的中心点作为Anchor吗? 6."....... where F is the feature map input......." 从VGG-16计算得来的特征映射图是如何计算的呢?这个特征映射图作为响应生成分支和回归分支的输入,请问这个特征大小是怎么设定的?

xieyunjiao avatar May 26 '20 00:05 xieyunjiao