Hello,
Usually, when speaking about the invariance of descriptors, one can assume that a descirptor is invariant so the distance between two features of different images under different transformation is equal zero. But, in reality, it is not the case. So, the question is: What is the distance that we can consider as a refernce to evalute the invariance of a descriptor.
the error in the difference is due to
you should define what type of noise you have and the error it injects in your descriptors and then add the tolerance from the descriptors.
well it is not an easy task. usually the invariance is demonstrated mathematically and in practice i set the threshold using ROC.
hope it helps. good luck.
Hello,
Generally, for descriptors based on Histogram of Oriented Gradient (HoG), Euclidean distance is used to evaluate the similarity. On the other hand, Hamming distance is calculated to perform the matching of Binary descriptors.
Good luck
It depends on the structure of descriptors that you use, where each distance is suitable between a special descriptors with respecting to the space where you apply your comparison.
Hi Mariem,
In practice, there are several factors that hinder arriving at the theoretical "0 distance" you are aiming for, a tolerance level has to be applied.
Having said that, I would advise you to furnish further details on the exact scenario. For instance, and as seen above, people may run into ambiguities concerning the two features (are they two different features?) of different images (different images or an image and its transformed version?) and different transformations (what kind?). If you supply an example, that would make it much easier for the RG members to address your question, otherwise we are let to guessing.
Thank you for your answers.
Abbas, an example is the benchmark Coil100 which contains 100 objects, for each one there are 72 images of different views. The transformation between two successive images is a rotation of 5°. An example of a descriptor that I used is Zernike moments so I apply it on those images.
Thank you Mariem for the clarification. Well, as Elie stated above, the norm is to prove the invariance mathematically. However, if that is not a feasible thing to do, I would suggest to resort to using empirical evaluations to demonstrate the invariance. For example, if you are using the ZMs to produce Zernike clusters then calculating the total classification rates (Zernike clusters vs original clusters) could give you a hint about the robustness of your extracted descriptor. But as you can see, you need a ground truth data set to compare it to. That was my immediate thought about the problem; hoping that other RG members have better suggestions.
Hope that helps!
Hello , zernike moments (modulus) are rotation invariant if each moment is normalized by the piston as the difference between two rotated features is a factor in the phase... many articles already proven this (with scale invariance to)
Bin, Y., Jia-Xiong, P., 2002. Invariance analysis of improved Zernike moments.
Now be careful those invariance in practice need some special precautions to be hold.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence