Xerox Corporation

Jorge Sanchez

Florent C Perronnin

Thomas Mensink

Yan Liu

Local descriptors are extracted from an image. An image vector is generated having vector elements indicative of parameters of mixture model components of a mixture model representing the extracted local descriptors. The image vector is compressed using a vector quantization algorithm to generate a compressed image vector. Optionally, the compressing comprises splitting the image vector into a plurality of sub-vectors each including at least two vector elements, compressing each sub-vector independently using the vector quantization algorithm, and concatenating the compressed sub-vectors to generate the compressed image vector. Optionally, each sub-vector includes only vector elements indicative of parameters of a single mixture model component, and any sparse sub-vector whose vector elements are indicative of parameters of a mixture model component that does not represent any of the extracted local descriptors is not compressed.

Image classification employing image vectors compressed using vector quantization

An input image representation is generated based on an aggregation of local descriptors extracted from an input image, and is adjusted by performing a power normalization, an Lp normalization such as an Lnormalization, or both. In some embodiments the generating comprises modeling the extracted local descriptors using a probabilistic model to generate the input image representation comprising probabilistic model component values for a set of probabilistic model components. In some such embodiments the probabilistic model comprises a Gaussian mixture model and the probabilistic model components comprise Gaussian components of the Gaussian mixture model. The generating may include partitioning the input image into a plurality of image partitions using a spatial pyramids partitioning model, extracting local descriptors, such as Fisher vectors, from the image partitions, and concatenating the local descriptors extracted from the image partitions.

Large scale image classification

A classifier training method and apparatus for training, a linear classifier trained by the method, and its use, are disclosed. In training the linear classifier, signatures for a set of training samples, such as images, in the form of multi-dimension vectors in a first multi-dimensional space, are converted to a second multi-dimension space, of the same or higher dimensionality than the first multi-dimension space, by applying a set of embedding functions, one for each dimension of the vector space. A linear classifier is trained in the second multi-dimension space. The linear classifier can approximate the accuracy of a non-linear classifier in the original space when predicting labels for new samples, but with lower computation cost in the learning phase.

Training a classifier by dimension-wise embedding of training data

A classifier method comprises: projecting a set of training vectors in a vector space to a comparison space defined by a set of reference vectors using a comparison function to generate a corresponding set of projected training vectors in the comparison space; training a linear classifier on the set of projected training vectors to generate a trained linear classifier operative in the comparison space; and transforming the trained linear classifier operative in the comparison space into a trained nonlinear classifier that is operative in the vector space to classify an input vector.

Growing community of inventors

Cordoba, Argentina

Jorge Sanchez