Classic paper Study bag of feature (ii)

Bag-of-word

Bag-of-words model is a common document representation method in information retrieval field. In the information retrieval, the bow model assumes that for a document, ignoring its word order and grammatical, syntactic and other elements, it is only regarded as a set of several words, each word in the document appears independent, does not depend on whether other words appear. For example, there are two documents:

1:bob likes to play basketball, Jim likes too. 2:bob also likes to play football games.

Based on these two text documents, a single word is extracted and a dictionary is constructed:

Dictionary = {1: "Bob", 2. "Like", 3. "To", 4. "Play", 5. "Basketball", 6. "Also", 7. "Football", 8. "Games", 9. "Jim", 10. "Too"}.

This dictionary contains 10 different words, according to the dictionary, the number of words in the above two documents are counted, each document can be represented as a 10-dimensional vector. As follows:

1:[1, 2, 1, 1, 1, 0, 0, 0, 1, 1] 2:[1, 1, 1, 1, 0, 1, 1, 1, 0, 0]

If the histogram statistics of the words in each type of document show a specific law, you can use this law to classify the mass document.

Bag of Feature

1.1 [CVPR06] Beyond Bags of features:spatial Pyramid Matching for recognizing Natural Scene Categories

absrtact : The idea of bow is introduced into the image, and Word replaces it with a particular feature descriptor in the image, but this completely ignores the spatial layout of the image, incapable of capturing shape or of segmenting an Object from its background, so it is implemented with spatial pyramid matching. Our method involves repeatedly subdividing the image and computing histograms of local features at increasingly fine Resol Utions.

comparison : The following three aspects are compared with traditional methods:

1Locally orderless IMAGES:SPM asAn alternative formulation of a locally orderless image, instead of a Gaussian scale space of apertures,define afixedhierarchy of rectangular windows.2 Multiresolution histograms:fixing The resolution at which the features is computed, but varying the spatial resolutio n at which they is aggregated.3Subdivide and Disorder:the best subdivision scheme is achieved when multiple resolutions is combinedincha principled; The reason forThe empirical success of "subdivide and disorder" techniques isThe fact that they actually performapproximate geometric matching.

Pyramid Match Kernels:

XY indicates that two vector pyramidmatch are used to calculate the appriosimate correspondence between XY. Placing a sequence of increasingly coarser grids over the FEA Ture space and taking a weighted sum of the number of matches that occur at each level of resolution.

Match means they fall into the same cell. Resolution counts from 0 to L.

At level L, image can is divide into 2exp (d*l) cells (the cell here should be the cluster center at the back?). ); The number of matchs at level L was given by (1);

The weight number of level L are set to 1/(2exp (L-l)),Note lower (coarser) level include the num of finer level, so The NUM of level L was given by. The Pyramid match kernel can be given by (2);

(1) (2)

(1) Why is the appriosimate correspondence between XY represented? How to calculate H, how to be considered fall into the same cell?

Spatial Matching Scheme

Perform pyramid matching in the two-dimensional image space, and use traditional clustering techniques in feature space. (for feature space in the image, the coordinates of the image already contain the geometric space information, only the vector can be arranged in the order of coordinates) (Feature space clustering will feature into the M-class channel, which is probably the above-mentioned fall into the same cell,h is the histogram statistics, I the smaller the correlation between the less)

The final dimension is:(in the above-described K (x, Y) is not added but each level of I vector connection into a very long vector); m=400 l=3 d=34000 is long and sparse.

Normalize All histograms is weight of all features in the image.

Q:

Local and global feature representations: This article says that SPM is a approximate global geometric correspondence, and how to understand an alternative formulation of a locally orderless Image, what is the definition of traditional local and global features, and what are these??

Ps:

Partly from: Http://blog.csdn.net/v_JULY_v

Classic paper Study bag of feature (ii)