Heyyy ,,,我要做分类。对于描述符,我计划使用LOWE中的HOG和SIFT描述符。
1. For HOG, is that true that we need to compute the gradient of all image's pixels??
For example we have image with size 10x10 pixels. And we compute the HOG to seek the orientation
for every pixels of image. And eventually we will get 100 orientations and generate histogram
(represent all of those image's pixels orientation). This Histogram is going to be used for the
classification??
And to get the scale invariant we need to perform this descriptor of different size (scale)
images??
SIFT是Scale Invariant Features Transform。所以它是比例和旋转不变 我读到from here在SIFT中我们需要使用高斯来平滑我们的图像 分辨率变低..
2. Why we have to do that?
And for scale invariant features, how to obtain that in SIFT?? Do we need to rescale our image
in every octave and then apply Gaussian filter in this new scaled images?? Or it's enough to
get scale invariant only in 1 octave by applying 3 times gaussian filter??
How about the histogram, Is it same with HOG that we have to compute all the pixels???
Thankss
答案 0 :(得分:2)
如果你计划至少实施SIFT,你应该阅读Lowe的论文,因为它是简历史上被引用次数最多的一篇!
在SIFT中,应用高斯平滑以计算DOG(高斯差)。然后执行Scale Extrema Detection,您将检测到特征点。
获得此功能点后,您需要为每个功能计算HOG。您无需为整个图像计算它!由于我们采用16x16邻域,因此结果将是128长度描述符。
比例不变是因为相对于键盘的比例,在每个关键点周围的窗口中对渐变进行采样。
但是,你应该再次阅读Lowe的论文,因为那里非常清楚。您还应该清楚一些概念,例如特征点,HOG,DOG等,以真正理解SIFT