"CV Knowledge Learning" Fisher Vector

In the paper "Action recognition with improved trajectories" see Fisher Vector, so learn. But a lot of information on the internet I think all write bad, check again, according to their own understanding of the statement, Hope Daniel Correct.

Kernel functions:

Let's take a look at the concept of the nuclear function described in the statistical learning method,

As can be seen, the kernel function is actually an inner product, in the SVM formula can be extracted from the inner product part. The data may be linearly irreducible in the low-dimensional input space, while the Gauveshilbert space may be linearly divided, thus passing through a mapping function. In fact, the inner product can be understood as similarity or distance.

Fisher Core:

Fisher Core is related to Fisher information matrix, the meaning of Fisher information Matrix, the answer on the understanding of http://www.zhihu.com/question/26561604.

Finally, the Fisher nucleus is related to the Fisher Information Matrix and gets: And I is the information matrix, and U is the score Function.

Application of Fisher vector in images:

The image is described here using GMM, and the description of the image acquisition is in the global scope, so the Fisher vector is finally described as a global feature. The general steps are as follows:

The data set is randomly selected to estimate the parameters of GMM. The Fisher vector should be the aggregation of the score function, but it will be multiplied by the Fisher information Matrix because it will be applied to the kernel function.

The GMM model is:

Which (people familiar with GMM are easy to understand),

According to the Bayesian formula, the probability of defining a descriptor that belongs to the I-Gaussian model is:

Score functions for each parameter

The approximate solution of Fisher information matrix is:

The FV vectors are calculated as:

The end result is:

It is important to note that the parameters have been estimated in the first step. The advantage of the FV vector is that it transforms each size of a different description subset into a consistent, uniform feature vector representation.

The algorithm from the original feature to the FV vector is expressed as follows:

Actually also did not think of things, originally thought last night to understand very good, this morning again to think about, not ah ... Directly to know the written on it, the amount, a lot of blog did not write to the point, hey ~~~~~ will be good ~~~~~~~~~~~ t_t

"CV Knowledge Learning" Fisher Vector