Vector transformation for indexing, similarity search and classification

A non-transitory computer-readable storage medium having executable computer program instructions embodied therein for boosting a video classification score, the actions of the computer program instructions comprising:

• retrieving a feature vector, the feature vector representing a target media object;
• generating a permutation, the permutation comprising a vector of ordinals, wherein the dimensionality of the permutation vector is equivalent to the dimensionality of the feature vector;
• permuting the feature vector with the generated permutation;
• truncating the permuted feature vector according to a selected window size;
• identifying the index of the maximum value of the truncated permuted feature vector;
• and
• producing a sparse binary vector, the sparse binary vector comprising the identified index encoded using one-hot encoding.

A computer system for encoding a feature vector, the system comprising:

• a computer processor;
• and
• a non-transitory computer-readable storage medium storing executable computer program instructions comprising:
• retrieving a feature vector, the feature vector representing a target media object;
• generating a permutation, the permutation comprising a vector of ordinals, wherein the dimensionality of the permutation vector is equivalent to the dimensionality of the feature vector;
• permuting the feature vector with the generated permutation;
• truncating the permuted feature vector according to a selected window size;
• identifying the index of the maximum value of the truncated permuted feature vector;
• and
• producing a sparse binary vector, the sparse binary vector comprising the identified index encoded using one-hot encoding.

A computer implemented method of encoding a feature vector, the method comprising:

• retrieving a feature vector, the feature vector representing a target object;
• generating a set of permutations, the permutations comprising vectors of ordinals, wherein the dimensionalities of the permutation vectors are equivalent to the dimensionality of the feature vector;
• permuting the feature vector with the set of permutations, creating a set of permuted feature vectors;
• truncating the set of permuted feature vectors according to a selected window size, creating a set of truncated permuted feature vectors;
• identifying the index of the maximum value of each truncated permuted feature vector in the set of truncated permuted feature vectors;
• producing a set of sparse binary sub-vectors, each sparse binary sub-vector comprising an encoded identified index of a truncated permuted feature vector;
• and
• concatenating the sparse binary sub-vectors to produce a sparse binary vector.