Introduction to the Hessian feature detector for finding blobs in an image


In many computer vision applications, it is important to extract special features from the image which are distinctive, unambiguously to locate and occur in different images showing the same scene. There is a complete subbranch in the computer vision field dedicated to this task: feature matching. Usually, this process consists of three tasks: detection, description and matching. I shortly want to summarize the steps but you can easily find further information on the web or in related books1.

  1. The first step is about the detection of distinctive points in an image. This could be a corner of an object, an intensity blob (e.g. the inner parts of the eyes) or any other special shape or structure you can imagine. Not good candidate points are e.g. straight lines. It is hard to detect a point on a line unambiguously in two images (which point on the line to choose?).
  2. The goal of the description step is to describe the area around a detected point. A lot of techniques can be used here and it is quite common to get a vector with some numbers which describe the area around the point as output. If two points have a similar surrounding, it would be good if the vectors would also have similar numbers.
  3. This is essential for the third and last step: matching. Imagine two images showing the same scene with a different viewpoint. In both images are distinctive points detected and the surrounding of each point described. Which points correspond, i.e. which points show the same part of the scene? This can e.g. be done by measuring the similarity between the description vectors.

For each of these steps, many different algorithms exist. As you may have guessed from the title of this article I want to focus here on the detection step. More precisely, I want to give an introduction to the Hessian detector which is mathematically defined as

\begin{equation} \label{eq:HessianDetector_Definition} \sigma^4 \cdot \det(H_{\sigma}) = \sigma^4 \cdot \left( \partial_{xx} \cdot \partial_{yy} - \partial_{xy}^2 \right) \end{equation}

with \(H_{\sigma}\) as the Hessian matrix calculated at the scale level \(\sigma\) and e.g. \(\partial_{xx} = \frac{\partial^2 I}{\partial^2 x^2}\) denoting the second order derivative of the image \(I\). I just wanted to give the definition of the detector at the beginning; the intention of this article is to give some insights why this equation might indeed be useful. Also, note that the detector is designed to detect blobs in an image (and not corners).

So, what exactly is a blob in an image? It is a homogeneous area with roughly equal intensity values compared to the surrounding. The ideal artificial example is a 2D Gaussian function where the intensity values equally decay in a circle way blending together with the surrounding – visible in the image as a smooth blob. A more realistic example would be the inner part of a sunflower which is usually dark compared to the bright leaves. The following figure shows examples for both. A blob is not necessarily related to a circle like structure though. Any kind of intensity notch may also be detected as a blob. And the idea is that these circles and notches are part of the object and therefore are also visible in other images showing the same object.

3D and density plot showing a Gaussian blob 17180 detected blobs in an image showing sunflowers
Left: blob of the Gaussian function \(G_{\sigma}(x,y) = \frac{1}{2 \pi \sigma^2} \cdot e^{-\frac{x^2+y^2}{2 \sigma ^2}}\) as 3D and density plot. Right: 17180 detected keypoints on a sunflower image2 using a blob detector (here the AKAZE features which use the Hessian detector were used). The circle size corresponds to the scale level in which the keypoint was detected (size of the object the keypoint covers) and the intensity of the circle colour roughly corresponds to the strength of the detected keypoint (more precisely: on the determinant response of the Hessian, but more on this later). The detector did not find the large sunflower blobs in the foreground but it worked quite well for sunflowers more away from the camera. Notice how the inner parts of the sunflowers relate to very strong keypoints (blobs).

The rest of this article is structured as follows: I begin introducing the Hessian matrix, analyse the curvature information it contains and describes how this information is used in the Hessian detector. But I also want to talk a bit about how the detector incorporates in the scale space usually used in feature matching to achieve scale invariance.

The Hessian matrix

Let's start with the Hessian matrix \(H\) (scale index \(\sigma\) for simplicity omitted). Mathematically, this matrix is defined to hold the information about every possible second order derivative (here shown in the 2D case)3:

\begin{equation*} H = \begin{pmatrix} \frac{\partial^2 I}{\partial^2 x^2} & \frac{\partial^2 I}{\partial x \partial y} \\ \frac{\partial^2 I}{\partial y \partial x} & \frac{\partial^2 I}{\partial^2 y^2} \\ \end{pmatrix}. \end{equation*}

Since \(\frac{\partial^2 I}{\partial x \partial y} = \frac{\partial^2 I}{\partial y \partial x}\) this matrix is symmetric4. Conceptually, \(H\) can be understood as the second order derivative of a higher dimensional function. As we know from the 1D case, the second order derivative tells us something about the curvature of a function: is it convex, concave or neither of those? The result here is a number where the signum function tells us which of the cases is true. But what is the curvature of a 2D function?

It turns out that things get more complicated when introducing another dimension (who on earth would have guessed that...). We can't express the curvature as a single number anymore since the question now is: how is the curvature in a certain direction? Imagine you are walking over an arête in the mountains where it goes down on the left and right side and straight on in front of you. What is the curvature when you keep moving ahead? Since there are not many changes in height probably (roughly) zero. And what if you move right directly downwards the precipice? Probably very high since there are a lot of changes in height. So, direction obviously matters.

To get the curvature of a 2D function we need to provide three parameters: the Hessian matrix \(H\), which acts as the curvature tensor containing the information about every direction, a specific direction \(\vec{s}\) in which we are interested and the location \(\left( x_0, y_0 \right)^T\) where we want to know the curvature of. The direction is usually provided as uniform (i.e. \(\left\| \vec{s} \right\|=1\)) column vector. We can then compute the so-called directional derivative

\begin{equation} \label{eq:HessianDetector_DirectionalDerivative} \frac{\partial^2 I}{\partial^2 \vec{s}^2} = \vec{s}^T H \vec{s} \end{equation}

which gives as the curvature (the second order derivative) in the direction \(\vec{s}\). Before showing some visualization, I want to provide a numeric example. For this, consider the following 2D function and suppose you want to know the curvature in the \(45^{\circ}=\frac{\pi}{4}\) direction at the origin

\begin{equation*} f(x,y) = \cos(x), \quad \vec{s} = \begin{pmatrix} \cos\left( \frac{\pi}{4} \right) \\ \sin\left( \frac{\pi}{4} \right) \end{pmatrix} \quad \text{and} \quad \begin{pmatrix} x_0 \\ y_0 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix}. \end{equation*}

It is actually a bit stupid to call this a 2D function since it is just the standard cosine but the resulting graph fits nice to my arête analogy. The first thing to do is to calculate the derivatives of the Hessian matrix

\begin{align*} \frac{\partial^2 f(x, y)}{\partial^2 x^2} &= - \frac{\partial}{\partial x} \sin(x) = -\cos(x) \\ \frac{\partial^2 f(x, y)}{\partial x \partial y} &= 0 \\ \frac{\partial^2 f(x, y)}{\partial y \partial x} &= - \frac{\partial}{\partial y} \sin(x) = 0 \\ \frac{\partial^2 f(x, y)}{\partial x \partial y} &= 0 \end{align*}

and summarize the results in the Hessian matrix

\begin{equation*} H = \begin{pmatrix} -\cos(x) & 0 \\ 0 & 0 \\ \end{pmatrix}. \end{equation*}

Next, we apply \eqref{eq:HessianDetector_DirectionalDerivative} and get

\begin{align*} \frac{\partial^2 f}{\partial^2 \vec{s}^2} = \vec{s}^T H \vec{s} &= \begin{pmatrix} \cos\left( \frac{\pi}{4} \right) & \sin\left( \frac{\pi}{4} \right) \end{pmatrix} \begin{pmatrix} -\cos(x) & 0 \\ 0 & 0 \\ \end{pmatrix} \begin{pmatrix} \cos\left( \frac{\pi}{4} \right) \\ \sin\left( \frac{\pi}{4} \right) \end{pmatrix} \\ &= \begin{pmatrix} \cos\left( \frac{\pi}{4} \right) & \sin\left( \frac{\pi}{4} \right) \end{pmatrix} \begin{pmatrix} -\cos(x) \cos\left( \frac{\pi}{4} \right) \\ 0 \end{pmatrix} \\ &= - \cos\left( \frac{\pi}{4} \right) \cos(x) \cos\left( \frac{\pi}{4} \right) \\ &= - \frac{1}{\sqrt{2}} \frac{1}{\sqrt{2}} \cos(x) \\ &= - \frac{\cos(x)}{2} \end{align*}

i.e. a curvature of \(-\frac{1}{2}\) in the direction of \(\vec{s}\) at our evaluation point \(x_0=0\) (origin).

You can imagine this as follows: at every point of our 2D function, we have a normal vector pointing orthogonal to the surface at that point. The normal vector forms together with the direction \(\vec{s}\) a plane in the 3D space. This plane intersects with our function resulting in a projection which is a 1D function and \eqref{eq:HessianDetector_DirectionalDerivative} is then just the curvature of this function (like from every other 1D function). The following animation shows the discussed points and lets you choose an arbitrary direction.


Animation which shows the directional derivative from \eqref{eq:HessianDetector_DirectionalDerivative} on the function \(f(x)=\cos(x)\). The slider controls the direction for which the curvature should be calculated (in degree). Besides the current direction \(\vec{s}\) (red) also the direction of the maximum curvature \(\vec{e}_1\) (blue) is shown (\(\vec{e}_1\) denotes the first eigenvector, but I come to this point later). The length of the vectors corresponds roughly to the magnitude of the current curvature.

Even though this is very exciting, one is often not interested in the curvature of an arbitrary direction. What is more from interest is the direction of highest and lowest curvature. The good news is that it is also possible to retrieve this information from \(H\). The information lies in the eigenvectors and eigenvalues: the eigenvector \(\vec{e}_1\) points in the direction of the highest curvature with the magnitude \(\lambda_1\). Similarly, \(\vec{e}_2\) corresponds to the direction of lowest curvature with the strength of \(\lambda_2\)5. As you may have noticed, I already included \(\vec{e}_1\) in the previous animation. Can you guess in which direction \(\vec{e}_2\) points and what the value of \(\lambda_2\) is?6

For a 2D function, this means that we can obtain two orthogonal vectors pointing in the highest and lowest direction. The following animation has the aim to visualize this. In order to make things a little bit more interesting, I also used a different function: \(L(x,y)\). This is now a proper 2D function, meaning that both variables are actually used. It is essentially a linear combination of Gaussian functions building a landscape; therefore, I called it Gaussian landscape (if you are interested in the exact definition, check out the corresponding Mathematica notebook, but since the function values do not really matter here I did not want to confuse you).

Curvature analysis in a Gaussian landscape. Two vectors are shown at the current position: the blue vector indicates the first eigenvector \(\vec{e}_1\) and points in the direction of highest curvature (at that point). Likewise, the second eigenvector \(\vec{e}_2\) denotes the direction of lowest curvature. Both vectors are scaled to roughly represent the magnitude of curvature (also shown precisely over the image in form of the eigenvalues \(\lambda_1\) and \(\lambda_2\)). You can use the top 2D slider to choose a different position.

Take your time and explore the landscape a bit. A few cases I want point to:

  1. If you are at the centre of the Gaussian blob (bottom left corner, \(\left( -10, -10 \right)^T\), did you notice that both eigenvalues are equal (\(\lambda_1 = \lambda_2\))? This is because the curvature is the same in every direction so there is no unique minimal or maximal direction. This is for example also true for every point on a sphere.
  2. When you are on the ground, e.g. at position \(\left( -6, -6\right)^T\), there is no curvature at all (\(\lambda_1, \lambda_2 \approx 0\)). This is true for every point on a planar surface.
  3. Another fact is already known from the previous animation: when you walk alongside the arête (e.g. at position \(\left( 0, -5\right)^T\)), the magnitude of the first eigenvalue is high whereas the second is roughly zero (\( \left| \lambda_1 \right| > 0, \lambda_2 \approx 0\)). This is for example also true for tube-like objects (e.g. a cylinder).
  4. Last but not least, when you descent from the hill (e.g. at position \(\left( 0, -9\right)^T\)) there is curvature present in both directions but the first direction is still stronger (\( \left| \lambda_1 \right| > \left| \lambda_2 \right| > 0\)). This is a very interesting point, as we will see later.

What has this to do with blobs in an image? Well, the first and last case I have just noted to fall directly into the definition of a blob. Whereas the first one is more circle-like and the last one is more notch-like. Since we want to detect both cases, it does not really matter which one we have. One way to detect these cases based on the eigenvalues is by using the Gaussian curvature

\begin{equation} \label{eq:HessianDetector_GaussianCurvature} K = \lambda_1 \cdot \lambda_2. \end{equation}

The main observation is that this product is only large when both eigenvalues are large. This is true for the first and fourth case and false for the other two since the product is already low when \(\lambda_2 \approx 0\) (note that it is assumed that the eigenvalues are descendingly sorted, i.e. \(\lambda_1 \geq \lambda_2\)).

We can now detect blobs at each image position by calculating the Hessian matrix via image derivatives, their eigenvalues and then the Gaussian curvature \(K\). Wherever \(K\) is high we can label the corresponding pixel position as a blob. However, wherever something is high is not a very precise formulation, and, usually, comes with the cost of a threshold in practice. Since we have also a strength of the blob (the determinant response) we can easily overcome this by using only the largest blobs in an image and these are the local maxima of \(K\). It is e.g. possible to sort the local maxima descendingly and use the \(n\) best (or define a threshold, or use just every maximum, or do something clever).

But wait a moment, when \eqref{eq:HessianDetector_GaussianCurvature} is already the answer to our initial problem (finding blobs) why does \eqref{eq:HessianDetector_Definition} use the determinant of \(H\) and this strange \(\sigma^4\) prefactor? The second part of this question is a bit more complicated and postponed to the next section. But the first part of this question is easily answered: there is no difference, since the equation

\begin{equation*} \det(H) = \lambda_1 \cdot \lambda_2 \end{equation*}

holds. This means that \eqref{eq:HessianDetector_Definition} already calculates the Gaussian curvature. Before moving to the discussion of the scale parameter though, I want to show the result when the determinant of the Hessian is applied to the Gaussian landscape.

Determinant response of the Hessian applied to the Gaussian landscape
Determinant response of the Hessian applied to the Gaussian landscape \(L(x,y)\). Grey level intensity is used to encode the magnitude of the determinant response whereas darker areas stand for a very low response (absence of curvature) and brighter areas are symbolic for a high response (presence of curvature). Additionally, the local maxima are labelled in red. Note that every local maximum denotes a blob in the function.

As you can see, the determinant response clearly separates our landscape with the maxima identifying locations which are either circle-like (bottom left) or notch-like (rest) blobs. Next, I want to apply this technique to a real image.

Airport image from a bird's eye view
Airport image7 from a bird's eye view.
Determinant response of the airport image at scale level 2
Determinant response of the Hessian applied to the airport image at the scale level \(\sigma_{2}=4\). In this case already \eqref{eq:HessianDetector_Definition} was use, i.e. including the scale normalization parameter which will be discussed in the next section. The image derivatives are calculated via Gaussian derivatives (with \(\sigma=4\)). The image is scaled for better contrast and comparison as \(R_{\sigma_2}\) like defined in \eqref{eq:HessianDetector_ResponseImageScale} (next section).

Notice the small intensity blobs present in the original image which are very well captured by the determinant response (e.g. the small white dots in the bottom of the image). Even though we can now detect some relevant parts of the image, did you notice that other parts are not detected? Especially the larger blobs like the round structure in the top right corner of the image. This is where the scale normalization parameter \(\sigma\) gets important and is the concern of the next section.

Scale normalization

So, the first important observation is that blobs may occur in different sizes. This was already the case for the sunflower image at the beginning. Not all sunflowers have the same size, and, additionally, they project to different sizes on the camera depending on the distance to the camera. What we want to achieve is a so-called scale invariance: blobs should be detected regardless of the size in the image.

But how can we achieve scale invariance? The answer is to build a scale space, e.g. by blurring the image repeatedly with a Gaussian function. The result is a scale pyramid with a different intrinsic image resolution on each level. This resolution is denoted by a \(\sigma_i\) parameter per level \(i\) which corresponds to a Gaussian function \(G_{\sigma_i}\) used to create the level (via convolution, i.e. \(I*G_{\sigma_i}\)). With increasing scale level the \(\sigma_i\) value also increases and the image is blurred more intense.

Our extrema search is therefore extended by an additional dimension: we can now search for maxima in the determinant response not only spatially in the image but also across scale in different resolutions of the same image. At each position, we can consider our determinant response also as a function of scale and pick the scale level with the highest response. But before we can do that we first need to solve a problem.

The thing is that the Gaussian blurring flattens the intensities in the image out since in each blurring step the intensity values come closer together due to the weighting process. Per se this is not a problem and is precisely the job of Gaussian blurring: flatten out outliers to create a smooth transition between the values. But the weighting process also decreases the possible range of values. If we, for example, start with an image where the range of intensity values lies in \([10;20]\) then after the blurring this range must be smaller, e.g. \([12;17]\) (depending on the distribution of the values and the precise \(\sigma_i\) value). At the end, when we use a really high \(\sigma_i\) value we have only one flat surface left with exactly one intensity value remaining and this is the average intensity of the first image (this is because the convolution with a Gaussian converges to the average image value).

Ok, but why is this now a problem? The fact is that when the intensity range decreases over scale it is also very likely that the derivative values will decrease (fewer changes) and hence the determinant response decreases. But when the determinant responses tend to get lower in higher scale levels we introduce a systematic error while searching for extrema across scale. Responses from higher levels will always have the burden of being computed based on a smaller intensity range compared to the first scale levels of the pyramid. Note that this does not mean that a response value for a specific location strictly decreases over scale. It is still possible that locally a new blob structure is detected which produces (slightly) higher response values than the previous level. But is it correct and significant? To find out, we must find a compensation for the decreased intensity range and we need to compensate more as higher we climb in our scale pyramid.

The idea is to use the \(\sigma_i\) values for compensation. Since they increase with higher scales, they fit our requirements. And this is exactly the reason why \eqref{eq:HessianDetector_Definition} contains \(\sigma^4\) as prefactor8 which stretches our intensity range again to be (roughly) equal as before the blurring process (e.g. \([10.4;19.7]\) instead of \([12;17]\); don't expect exact results in a noisy and discrete world). This makes the determinant response \(\det(H_{\sigma})\) comparable across scale. With this problem now solved we can move on and detect extrema over scale. For this, consider the following figure which shows the determinant response for the position in the blob of the top right corner of the airport image (also labelled in the image below).

Determinant response over different scale levels with and without scale normalization
Determinant response over 50 scale levels beginning with \(\sigma_1=2\) and proceeding further in steps of \(\sigma=2\) up to \(\sigma_{50}=100\). The position of interest is \((x_0,y_0)=(532, 98)^T\) and the responses are calculated using Gaussian derivatives with the corresponding \(\sigma_i\) per scale level. The determinant response with (orange) and without (blue) the scale normalization prefactor is shown. The maximum for both cases is depicted with a circle. Note the different plot ranges for the two cases shown on the left respectively right side.

Technically, we have a unique maximum in both cases. Even when we do not normalize we get a peak. However, it is not as wide as in the normalized case. But back to the question: is this peak correct and significant? To answer the question let me emphasize that this is only the trend for a specific location and when you look at the original image you see that in a very fine scale space there is just a white area (i.e. roughly zero determinant response). But when increasing scale the borders of the circle-like blob get visible and we have therefore an increase in response. This doesn't answer the question about correctness and significance, though. For this, let's consider the following figure which shows the global response range for all response values in all locations over scale for both cases.

Range of determinant response values in all locations as a function of scale
Minimum (dotted) and maximum (solid) determinant response values in all locations as a function of scale. The range for both with (orange) and without (blue) scale normalization is shown. The maximum in both cases is depicted with a circle. Note again the different plot ranges for the two cases.

This graph shows very clearly what I meant with the decrease of response over scale when no normalization is used. In this case, we start with a large range in the first levels but the range narrows very fast to a nearly zero wide range. This explains why the magnitude of the response values have been so low in the previous figure (e.g. only about 0.00006 at the peak) and really reinforces doubts about the correctness and significance of the peak. However, the range of the normalized response gives a different picture. Even though the range does not stay fixed (but this wasn't expected and highly depends on the example image), it's width remains rather solid. No strict decay and the values do not get ridiculously small. This shows very clearly that scale normalization is indeed important.

To answer the question about correctness and significance (finally!), let's compare the maximum response from our specific location \((x_0,y_0)=(532, 98)^T\) with the global maximum response of all locations in all scale levels9

\begin{align*} \frac{ \max_{i=1,\ldots,50}(\det(H_{\sigma_i}(532, 98))) }{ \max_{i=1,\ldots,50}(\det(H_{\sigma_i})) } &= % MathJax offers a way to apply css to math elements: http://docs.mathjax.org/en/latest/tex.html#html % The math must contain a symbol, i.e. it should not be empty. Therefore, an arbitrary symbol is printed with the same colour as the face of the circle \frac{ \style{border: 2.2px solid #5e81b5; border-radius: 50%; background: black; width: 0.6em; height: 0.6em;}{\color{black}{o}} }{ \style{border: 2.2px solid #5e81b5; border-radius: 50%; background: white; width: 0.6em; height: 0.6em;}{\color{white}{o}} } = \frac{0.0000621447}{1.10636} = 0.0000561703 \\ \frac{ \max_{i=1,\ldots,50}(\sigma_i^4 \cdot \det(H_{\sigma_i}(532, 98))) }{ \max_{i=1,\ldots,50}(\sigma_i^4 \cdot \det(H_{\sigma_i})) } &= \frac{ \style{border: 2.2px solid #e19c24; border-radius: 50%; background: black; width: 0.6em; height: 0.6em;}{\color{black}{o}} }{ \style{border: 2.2px solid #e19c24; border-radius: 50%; background: white; width: 0.6em; height: 0.6em;}{\color{white}{o}} } = \frac{15.3165}{34.6083} = 0.442567. \end{align*}

The idea behind this is that we need to compare different response values over scale for extrema detection and this relation measures how near the values are to each other. As you can see, the maximum response without normalization is absolutely not significant and hence does not represent a correct extremum. On the other hand, the normalized response does not suffer from a huge difference between the values and therefore I would conclude this to be a very significant and correct extremum.

To finish up, let's take a look at the response image for the different scale levels. In order to make responses across scale visually comparable and to achieve a better image contrast, I scaled each response value with the minimum and maximum over all scale levels and all pixel locations, i.e. the response image \(R_{\sigma_i}\) for the scale size \(\sigma_i\) is calculated as

\begin{equation} \label{eq:HessianDetector_ResponseImageScale} R_{\sigma_i} = \frac{ \sigma_i^4 \cdot \det(H_{\sigma_i}) - \min_{i=1,\ldots,50}(\sigma_i^4 \cdot \det(H_{\sigma_i})) } { \max_{i=1,\ldots,50}(\sigma_i^4 \cdot \det(H_{\sigma_i})) - \min_{i=1,\ldots,50}(\sigma_i^4 \cdot \det(H_{\sigma_i})) }. \end{equation}

This way all response values stay in \([0;1]\) and a response value of e.g. 0.4 means the same in all images, i.e. does not depend on the current scale level. But this formula is only used to achieve a better visualization; feel free to ignore it.

Airport image from a bird's eye view
Airport image with the analysed position \((x_0,y_0)=(532, 98)^T\) labelled in red.

Determinant response over the complete scale space. The image is scaled for better contrast and comparison as \(R_{\sigma_i}\) like defined in \eqref{eq:HessianDetector_ResponseImageScale}.

As you can see, the blob at our location of interest \((x_0,y_0)=(532, 98)^T\) is now detected and \(\sigma_{16}=32\) gives indeed a fairly strong response. With the same technique we can also detect other blobs: find the maximum determinant response across scale for each image position.

This was also the last part of our journey which started with \eqref{eq:HessianDetector_Definition} were we took a deep breath at the Hessian matrix and analysed the curvature information it contains. We moved forward while keeping the Gaussian curvature \(K\) in our bag and used this technique to find extrema responses in the image. We needed to take a break and first got around about correct scale normalization. This problem solved via the \(\sigma_i^4\) prefactor, we could search for extrema over the complete scale pyramid and so the circle completed again with \eqref{eq:HessianDetector_Definition}.

List of attached files:

  • HessianDetector_GaussianBlob.nb [PDF] (Mathematica notebook which I used to create the Gaussian blob image)
  • HessianDetector_HessianCurvature.nb [PDF] (Mathematica notebook which contains the analysis of the curvature information embedded in the Hessian matrix, e.g. it contains the first two animations I used here)
  • HessianDetector_DeterminantResponse.nb [PDF] (Mathematica notebook for the calculation of the determinant response; both on the Gaussian landscape and on the airport image. Contains also the figures for the local response trend and global response range)

1. I can e.g. recommend chapter 16 (p. 493ff) of the Learning OpenCV 3 book by Gary Bradski and Adrian Kaehler. If you want a deeper understanding I would recommend looking at a specific feature detector. E.g. the book Digital Image Processing (2nd edition) by Burger and Burge has a great chapter about SIFT (chapter 25, p. 609ff).
2. Image from Frank Vincentz (Own work) [GFDL or CC-BY-SA-3.0], via Wikimedia Commons.
3. It is possible to derive the definition of the Hessian matrix but I will only present the results here. For more information see e.g. chapter 3 (p. 88ff) of the book Data Visualization - Principles and Practice (2nd edition) by Alexandru C. Telea.
4. Due to a property called symmetry of second derivatives.
5. If you want to know why this is true, you may are interested in reading chapter 7 (p. 254ff) of the book Data Visualization - Principles and Practice (2nd edition) by Alexandru C. Telea.
6. It is probably not too hard to see that the second eigenvector points in the \(\alpha = 90^{\circ}\) direction (the \(\alpha = 270^{\circ}\) direction has an equal curvature in this case), i.e. \(\vec{e}_2 = \begin{pmatrix} \cos(90^{\circ}) \\ \sin(90^{\circ}) \end{pmatrix} = \begin{pmatrix} 0 \\ 1 \end{pmatrix}\) with \(\lambda_2=0\) (zero curvature).
7. Part of an image by Hynek Moravec (Self-photographed) [GFDL, CC-BY-SA-3.0 or CC BY 2.5], via Wikimedia Commons
8. To be honest, I don't know why the exponent \(^4\) is used here. What I can guess is that it starts with \(\sigma\) and per derivative this factor gets somehow squared, so we have \(\sigma\) for \(\partial_x\) and \(\partial_y\) and \(\sigma^2\) for \(\partial_{xx}, \partial_{yy}\) and \(\partial_{xy}\). When we plug this into the determinant formula \begin{equation*} \sigma^2 \partial_{xx} \cdot \sigma^2 \partial_{yy} + \sigma^2 \partial_{xy} \cdot \sigma^2 \partial_{xy} = \sigma^4 \cdot \left( \partial_{xx} \cdot \partial_{yy} - \partial_{xy}^2 \right) = \sigma^4 \cdot \det(H_{\sigma}) \end{equation*} we get \eqref{eq:HessianDetector_Definition}. You can find more about this topic in the paper Feature Detection with Automatic Scale Selection by Tony Lindeberg. But the important concept here is that there must be some kind of compensation, otherwise, extrema detection would not work.
9. Values correspond with the circles shown in the figures. The exact values are calculated in the Mathematica notebook.