I'm trying to implement my own version in MATLAB of the SIFT algorithm. I know that there are a lot of existing implementations but I would like to develop my own version to better understand the whole process.
At the moment I'm able to find the keypoints in the different scale-space and octaves but I didn't understand how to bring back to the original octaves the keypoints that I found in the other octaves. Is it done with a simple interpolation?
I read the original article of Lowe and also other tutorials but I miss this point. Any suggestion is greatly appreciated!
Related
Problem:
Atmospheric turbulation problem occurs due to foze,temperature variations, looking at to long distances with camera. There is a problem sample below
Example:
What I tried so far:
When I searched, I found this article which suggests to use
Dual Tree Complex Wavelet Transform(DTCWT). I followed through the
article and it seems like a good way to apply. Then I found this
implementation of the article in python code. I have run it and
seen the results but not as expected well.
I have simply took the average of some frames(let say 10), and get a result. I did it because the shaking pixels will be mostly true ones at any time. My results seems not bad after averaging.
What I need:
After searching much I am sure that this solution needs more frame to get one good frame. To do that I need a good image fusion algorithm.
I kindly request that you can suggest me any algorithm or any approaches I can try and go throuh to solve problem.
Note: I am familiar with OpenCV and C++ so priority is on these.
I'd like to detect regions in an image which contain a comparatively large amount of small details, but equally I need to ignore strong edges. For example I would like to (approximately) identify regions of small text on a poster which is located on a building, but I also want to ignore the strong edges of the building itself.
I guess I'm probably looking for specific frequency bands, so approaches that spring to mind include: hand tuning a convolution kernel(s) until I hit what I need, use specific DCT coefficients, apply a histogram on directional filter responses. But perhaps I'm missing something more obvious?
To answer a question in the comments below, I'm developing in Matlab
I'm open to any suggestions for how to achieve this - thanks!
Here is something unscientific, but maybe not bad to get folks talking. I start with this image.
and use the excellent, free ImageMagick to divide it up into tiles 400x400 pixels, like this:
convert -crop 400x400 cinema.jpg tile%d.jpg
Now I measure the entropy of each tile, and sort by increasing entropy:
for f in tile*.jpg; do
convert $f -print '%[entropy] %f\n' null:
done | sort -n
and I get this outoput:
0.142574 tile0.jpg
0.316096 tile15.jpg
0.412495 tile9.jpg
0.482801 tile5.jpg
0.515268 tile4.jpg
0.534078 tile18.jpg
0.613911 tile12.jpg
0.629857 tile14.jpg
0.636475 tile11.jpg
0.689776 tile17.jpg
0.709307 tile10.jpg
0.710495 tile16.jpg
0.824499 tile6.jpg
0.826688 tile3.jpg
0.849991 tile8.jpg
0.851871 tile1.jpg
0.863232 tile13.jpg
0.917552 tile7.jpg
0.971176 tile2.jpg
So, if I look at the last 3 (i.e. those with the most entropy), I get:
The question itself is too broad for a non-paper worthy answer on my side. That being said, I can offer you some advice of narrowing the question down.
First off, go to Google Scholar and search for the keywords your work is revolved around. In your case, one of them would probably be edge detection.
Look through the most recent papers ( no more than 5 years ) for work that satisfies your needs. If you don't find anything, expand the search criteria or try different terms.
If you have something more specific, please edit your question and let me know.
Always remember to split the big question into smaller chunks and then split them into even smaller chunks, until you have a plate of delicious, manageable bites.
EDIT: From what I've gathered, you're interested in an edge detection and feature selection algorithm? Here are a couple of helpful links, which might prove useful:
-MATLAB feature detection
-MATLAB edge detection
Also this MATLAB edge detection write up, which is a part of their extensive guide documentation will hopefully prove useful enough for you to dig through the Matlab image processing toolbox. documentation for specific answers to your question.
You'll find Maximally Stable Extremal Regions (MSER) useful for this. You should be able to impose an area constraint to filter out large MSERs and then calculate a MSER density, for example as Mark had done in his answer by dividing the image into tiles.
Can someone explain briefly how SDM (Supervised Descent Method) for Feature Extraction works?
I searched a lot on the Internet but couldn't found what I was looking for.
Is it only for feature extraction in videos, or can it be used in both videos and images?
If someone can explain, it would be of great help.
SDM is a method to align shapes in images. It uses feature extractors (SIFT and HoG) in the process, but is not a feature extractor.
Similar methods are ASM, AAM or CLM, but SDM has better performance and accuracy.
In the case of SDM, in the training process, the system learns some descent vectors from a initial shape configuration (different from the shapes in the database) to the database sets. Those vectors have the hability of fitting a new initial configuration with the face shape in the image you want to fit in.
This link can help you to learn more about it: http://arxiv.org/pdf/1405.0601v1.pdf
About the code, there is some demo samples in the main page of IntraFace but if you are looking for the code, I don´t think you can find it.
You can use vl_sift for beginning its working even more precise then their original descriptor however its not fast as their descriptor for real time implementation.
Regarding to their implementation no code were released so far. They are using some specialized version with very fast histogram calculation.
I want to start learning about algorithms so I began reading The Algorithm Design Manual by Steven Skiena because it is recommended in some threads I read in SO. However, I just stopped here
and can't understand most of it because many explanations are represented in figures or images and my screen reader recognize but can't read them.
For example,
"The nearest neighbor rule is very efficient, for it looks at each pair of points
tex2html_wrap_inline23349 //That's how my screen reader read it, I asume it's an image.
at most twice, once when adding
tex2html_wrap_inline23351 //image
to the tour, the other when adding
tex2html_wrap_inline23353 //another image
Against all these positives there is only one problem. This algorithm is completely wrong."
This is really frustrating for me because I'm beginning to enjoy this though I can understand why those images help a lot of readers.
So is their a way to understand this thing without seeing the figures? Or should I read another book?
Thanks in advance and happy new year everyone.
Considering these algorithms are dealing with geometrical analysis, I am afraid it would be difficult to understand them without the images, and even more difficult to replace these images with an equivalent textual description.
I am a beginner in image processing. I want to write an application in C++ or in C# for
Searching an image in a list of images
Searching for a particular feature (for e.g. face) in a list of images.
Can anybody suggest where should I start from?
What all should I learn before doing this?
Where can I find the correct information regarding this?
In terms of the second one, you should start off with learning how to solve the decision problem of whether a square patch contains a face (or whatever kind of object you are interested in). For that, I suggest you study a little bit of machine learning, the AdaBoost algorithm, Haar features, and Viola-Jones.
Once you know how to do that, the trick is really just to take a sliding window across your image, feeding the contents of that window into your detector. Then you shrink your main input image and repeat the process until your input image has gotten smaller than the minimum size input for your detector. There are, of course, several clever ways to parallelize the computation and speed it up, but the binary detector is really the interesting part of the process.
You may find some of the material linked from the CSE 517: Machine Learning - Syllabus helpful in getting into machine learning and understanding AdaBoost. You will certainly find the Viola-Jones paper of interest.