image-processing - 哈里斯-拉普拉斯-检测器: Corner- and Blob-Detector?

Question

我对Mikolajczyk等人介绍的Harris-Laplacian-Detector有疑问。使用哈里斯(Harris)，您可以搜索每个特定刻度内的“拐角”最大值，然后使用拉普拉斯算子(Laplacian)，可以在比找到的哈里斯点的刻度大一小的刻度上搜索“起泡”的最大值。

为什么最大化“ Blob ”超出范围对角点有如此好的影响？我认为，最好是寻找最大的“拐角”(例如，在刻度上找到最大的哈里斯)以找到良好的拐角点。

Answer 1

需要包含拉普拉斯算子以确保尺度不变。

尽管采用比例尺的哈里斯探测器确实具有非常好的可重复性(就探测位置而言)，但是比例尺的选择仍然是一个问题。他们注意到

during our experiments we noticed that the adapted Harris function rarely attains maxima in 3D space. Therefore, we propose to use a different function, the Laplacian, for scale maxima detection.

(来自 Indexing based on scale invariant interest points)

这在 his thesis中有更详细的解释:

In our experiments (cf. section 3.2.4) we noticed that the scale adapted Harris function rarely attains maxima over scales in a scale-space representation. If too few interest points are detected, the image is not reliably represented. Therefore, we abandoned the idea of searching 3D maxima of the Harris function. Furthermore, the experiments showed that LoG function enables the highest percentage of correct characteristic scales to be found. Therefore, we propose to use the Laplacian to select the scales for points extracted with the Harris detector. Harris-Laplace detector uses the Harris function (cf. equation 4.1) to localize points in each level of the scale-space representation. Next, it selects the points, for which the Laplacian-of-Gaussian (cf. equation 4.2) attains a maximum over scale. In this way we combine these two methods to obtain a reliable interest point detector invariant to signicant scale changes.

对于哈里斯(Harris)函数为什么没有在缩放比例上给出很多最大值的问题，我没有任何直观的解释，但是从经验上讲，他们发现情况确实如此。似乎没有什么阻止您使用Harris比例尺最大空间，但是您可能只会得到更少的检测结果。