ammai

Scale & Affine Invariant Interest Point Detectors

In this paper it is extend scale-invariant feature transformation to affine invariant one.In introduction, the author briefly introduces the history of interesting point detection, and points out all systems can not deal with non-uniform affine changes. He tries to propose a system can solve this problem even without any prior knowledge. The system extends the existing Harris-Laplace approach to achieve this goal.Some other works are mentioned in Related work section. The author tries to argue why they choose these methods to design an affine-invariant system.
After that, the scale-invariant interesting point detector and affine-invariant one are introduced subsequently. In scale-invariant one, an iterative-tuning method and a simplified but efficient one are proposed. In affine-invariant one, second moment matrix are also included, which tries to find the the transformation that projects the anisotropic pattern to the isotropic one. It is adopted to solve the problem of non-uniform affine transformation. After many equations shown, a psudo-code of it are presented. Then LoG is used to attain a maximum over scale. Lowe’s work in DoG accelerate the computation process, so here, the Harris-Laplace algorithm also can be simplified to accelerate.

Distinctive Image Features from Scale-Invariant Keypoints

This is a famous paper.This paper present a method to detect the interest points which are invariant to scale or affine transformations. This lecture contained a very useful technique in information retrieval these years.

In this paper it proposes a approach method which named "Scale Invariant Feature Transform(SIFT)", transforms the image data into scale-invariant coordinates relative to local features.

Here are 4 major approach stages,

1. Scale-space extrema detection - uses the DoG function to find the potential interest points, by comparing the 26 neighbor points and be selected if the DoG is larger or smaller then all others.
2. Keypoint localization - a detailed model is fit to determine location and scale. To delete the candidate points which contrast is low or on the edge.
3. Orientation assignment - one or more orientations are assigned to each keypoint location based on local image gradient directions. And the image data are transformed by this orientation to make sure the invariant of rotation.
4. Keypoint descriptor - represents these points with the parameters of location, scale, and orientation which assigned to these points.

and the end the keypoint descriptors generated by the above stages are highly distinctive, which is invariant in rotation, scale, and point of view

When we want to each detected point, a transformation matrix is also extracted which transforms the local patch around the detected point into a canonical form for efficient description and matching.
How to update the affine transformation is a querstion of this problem . The author uses the second moment matrix (co-variance) around the local patch and many issues are not discussed: the effect of sampling, window size, etc. Nevertheless, this paper still provides me a good example on the utilization of the continuous optimization technique.
Though I know this paper for a long time, I didn't attempt to read it before. This paper present a method to detect the interest points which are invariant to scale or affine transformations. That is, for each detected point, a transformation matrix is also extracted which transforms the local patch around the detected point into a canonical form for efficient description and matching.
Though the formulation and the method is elegant, it is too slow for many applications. The detection part requires around 50 times computation than SIFT, and the descriptor used in the paper is not as good as the one used in SIFT. I think that's why most applications only apply SIFT. It's interesting that some people use SIFT and this detector to detect the interest points and then use the SIFT descriptor to describe the detected points. These algorithms find different types of interest points and the union of these points cover a boarder class and improve the results. In some applications this is critical: SIFT sometimes give like 10 matchings between images.

local feature

When we talk about local feature, we also talk about local feature detector, scale invariant transform, RANSAC and SIFT tools.
Meansuring image/photo similarity by bags of local feature and some little region to find two images similarity. In local feature, we have a detector and a descriptor to descript the feature of an image. but here has some problems, for example: feature detecots what is the definition of a feature point? where are them in the image?
feature descripots: how to represent the feature in the image ?
So the classes of local detecor is ideally a good detecotr should recognized the corner point by looking through a small window (user define a sliding window). Then, a "flat region no change in all direction. (or change is very small). if we consider the location of the window region then the chager is large. So when we move the window along the edge direction the edge direction is no change.
The other problem, when the object scale is different each other. In this time, similar features (segments) of different scales should be detected respectively.
The scale invariant detection about its common solution we design a function on region size which is "scale invariant " for example :averaging intensity. So the point is take a local maximum of this function size. however how to fina a good finction for scale detection? It is a issue of this research. For example:Band-PASS flter
A method of this difference of gaussian use for scale invariance.
Difference Gaussian method has difference performance of target image. How to use feature values of image?
Interest point detection by difference of Gaussians (DoG) by x y scale.
y axis is the scale has difference of gaussian. local maxima or minima to its 26 neighbors. further filtering by constraints and other properties. two constrainst:
1. see the coordinate difference between two image. How about its constant?
if lower to the threshold then delete it
2. along the axis. Don't find all the candidate point in the same edge direction.
Rotationally invariant descriptors
local descriptors by image gradients, how to define which direction is more powerful, and has more information of this image.
We want to find a rule to detect the direction of the target information. So we can find a dominant direction of gradient and use it to be a baseline. Then computing image dericatives to this orientation.
So above all the SIFT (scale invariant feature transform), first we determining the scale (for example: by DoG .....etc.). Second using dominant gradient direction for local orientation and for invariance to scale and rotation, then computing gradient orientation histograms of several small windows (128 values for each point )
finally,normalize the descripter to make it incariant to intensity change. Then remember the 32 dimensions values to be a feature point of an image region.
RANSAC (Random sample consensus)
local features, measuring image/photo similarities by bags of local feature
for example: comparing two photos by local feature (RANSAC)
removing outlierss by spatial verification (RANSAC), assuming the outliers do not comply with the global affine transform, then cn be a search.
In RANSAC method estimating parmeter of a mathematical model by random sampling. Then a set of observed data which contains inliers and outliers.
About outlier, data points which do not fit the model. BUT, here are some disadvantages of this method.

Image Retrieval: Ideas, Influences, and Trends of the New Age

The process of retrieving and displaying relevant images based on user’s queries from a database. The contributing factors for image search are increase in a availability and demand of digital images and decrease in costs for storage capacity and processing.
The methods of them are text-based image retrieval TBIR, content-based image retrieval CBIR. The feature is color, shape, texture, spatial location…etc.
When begin an image query-simple visual feature query
-feature combination query, the specify values for combination of different features.
-localized feature query, the user indicates feature values and locations by placing regions on canvas.
-query by example:

How to give a good research talk

Research is communication from people to people. The greatest ideas are worthless if you keep them to yourself. So we should communicate them to others and get some feedback and relationship from the listeners.
A good papers and talks are a fundamental part of research excellence. When we giving a good talk, the presentation is about how to give a good research talk what your talk is for , what to put in it , how to present it.
When we make a presentation, first we focus on motivation, second focus on our key idea.
Motivation: have two minutes to engage the audiences before they start to doze
Key idea: must identify a key idea; leave the audience to figure it out for themselves.
Do not present related work: but must know the related work, respond readily to questions and do not disparage the opposition.
Write the slides the night before (or at least polish it then) : the talk absolutely must be fresh in mind.
Handwritten slides are fine, use permanent ink and gen an eraser.
By far the most important thing is to be enthusiastic.
When presenting your slides a very annoying technique is to reveal.