Hi you are absolutely right when you think that your results are wrong, because you are comparing the pixel intensity only. You need an algorithm to detect objects into the image, made classes of objects and then compare both, you need to study pattern recognition area specifically image processing in pattern recognition, books like Introduction to the pattern recognition and Enfoque lógico combinatorio can give you an idea. I have both in spanish language.

another possibility you have is search in internet or in Google scholar for algorithms to do this task.

Best regards

Hi Prashant,

Detecting common objects from only two images is a state-of-the-art research topic in computer vision, known as co-segmentation. This is not an easy task.

http://ivipc.uestc.edu.cn/project/FanmanMeng_SMCB/

From the above link, you can find one of the most recent papers on co-segmentation, and a MATLAB code of it.

If you prefer simpler solutions, maybe you can follow the following steps: 1) interest point detector, e.g. Harris, Hessian, or MSER, 2) local feature descriptor, e.g. SIFT, 3) represent the images by a set of local features, and 4) compare the local features across the images and find the common local features.

You may notice that even this SIMPLE solution will be very time-consuming to implement. I suggest you first try the MATLAB code described above.

Thank you for your answers.

Actually i have a video and i am tying to detect objects in it

So i converted that video in number of frames and then trying to compare each frames

With its next frame to detect object. Is it right approach? Or i should change the approach?

Firstly i dont have not done any work in image processing so please suggest some proper way for object detection.

Hi Prashant,

If you are going to detect an object from a video, it becomes another problem, known as object detection and tracking. I do not think co-segmentation can solve your problem.

An example is face tracking, in which face is detected from the first frame and then tracked over the remaining frames. Face tracking may be the easiest example because there have been a lot of automatic face detector in this field.

If the object is hard to be automatically detected, people usually let the user to specify the object in the first frame, and then let the object tracking algorithm to automatically track the object. Maybe this is your case.

Anyway, this is not an easy task that can be finished in one or two weeks. You have to read related articles on object tracking. Related keywords can be like "Kalman filter", "particle filter", etc.

I actually am not familiar with object tracking. You should ask a new question on source codes of object tracking. I know a C++ library known as OpenCV containing tracking algorithms, but the installation of this library is very complicated.

Good luck.

Hi Prashant, have you tried to use ASIFT for object detection. ASIFT (Affine Scale Invariant Feature Transform) based method for object detection shows good performance. Here the link for the paper: http://www.ipol.im/pub/art/2011/my-asift/ASIFT_SIIMS_final.pdf

Here is the link for a c++ implementation of ASIFT.

I hope that this will be helpful for you.

Hi Alessandro,

ASIFT is a very interesting affine-invariant local feature descriptor that can be useful for object retrieval and recognition. But for Prashant's task, I think object tracking techniques may be more helpful.

BTW, could you please show the link of the C++ implementation of ASIFT? I personally am quite interested in this feature descriptor.

Me too, I would love to see the C++ implementation of ASIFT Alessandro

Hi,

you can use SIFT algorithm to extract the key points from your image and then generate features descriptors.

To find all matching points between the two images you need to use Homography matrix with RANSAC algorithm.

There are plenty of MATLAB implementations of these algorithms on the internet.

I used these algorithms in image processing project last semester about (Image Mosaicing), I don't know if your approach is the right approach for object tracking in video , you need to do more search on the internet or communicate with someone who knows.

I hope this helps you and I recommend using MATLAB rather than c++ when writing codes for image processing because it is much easier and there are extensive set of functions for processing images.

Does anybody know how to get pixel images from SIFT descriptor in MATLAB? Like bag-of-feature , i want to visualize the patterns in the bag. I searched for a long time and found the same questions ask by other researchers, but no answer is provided.

One simply method represents the correlation.