Telecom Italia S.p.a.

Telecom Italia Lab S.p.a.

Gianluca Francini

Skjalg Lepsoy

Andrea Varesio

Simone Bizzarri

Giorgio Ghinamo

Massimo Balestri

Mauro Quaglia

Fabrizio Antonelli

Giuseppe Minerva

Giovanni Cordara

Claudio Lande

Marco Vecchietti

Nicola Reale

Fabrizio Moggio

Nicola Adami

Riccardo Leonardi

Attilio Fiandrotti

Giovanni Presti

Enzo Tartaglione

Pedro Porto Buarque De Gusmao

Diego Gibellino

Sergio Benini

Skjalg LEPSØY

Pedro Porto Buarque De GUSMÃO

Skjalg Telecom Italia SpA Lepsoy

Buarque De Gusmao Pedro Porto

A method includes providing a neural network having a set of weights. The neural network receives an input data structure for generating a corresponding output array according to values of the set of weights. The neural network is trained to obtain a trained neural network. The training includes setting values of the set of weights with a gradient descent algorithm which exploits a cost function including a loss term and a regularization term. The trained neural network is deployed on a device through a communication network, and used by the device. The regularization term is based on a rate of change of elements of the output array caused by variations of the set of weights values.

Neural networks having reduced number of parameters

A method in which a convolutional neural network is configured to receive an input data structure including a group of values corresponding to signal samples and to generate a corresponding classification output indicative of a selected one among plural predefined classes. The convolutional neural network includes an ordered sequence of layers, each configured to receive a corresponding layer input data structure including a group of input values, and generate a corresponding layer output data structure including a group of output values by convolving the layer input data structure with at least one corresponding filter including a corresponding group of weights. The layer input data structure of the first layer of the sequence corresponds to the input data structure. The layer input data structure of a generic layer of the sequence different from the first layer corresponds to the layer output data structure generated by a previous layer in the sequence.

Convolutional neural networks, particularly for image analysis

A monitoring system assesses the position of at least one target within an area. The monitoring system comprises a processing unit adapted to receive, from a camera, at least a frame depicting a scene of the area. The frames generated by the camera have a first resolution. The processing unit comprises: a database module storing, for each at least one target, a corresponding reference image, each reference image having a second resolution lower than the first resolution; a segmentation module configured to segment a received frame into frame segments, each frame segment having a resolution corresponding to the second resolution; a search module configured to implement a visual search procedure between each frame segment and a reference image comprising a depiction of the at least one target; and a position and translation module for calculating the position of the at least one target within the received frame.

Method and system for identifying targets in scenes shot by a camera

A method () for comparing a first video shot (Vs) comprising a first set of first images (I(s)) with a second video shot (Vs) comprising a second set of second images (I(t)), at least one between the first and the second set comprising at least two images. The method comprises pairing () each first image of the first set with each second image of the second set to form a plurality of images pairs (IP(m)), and, for each image pair, carrying out the operations a)-g): a) identifying () first interest points in the first image and second interest points in the second image; b) associating () first interest points with corresponding second interest points in order to form corresponding interest point matches; c) for each pair of first interest points, calculating () the distance therebetween for obtaining a corresponding first length; d) for each pair of second interest points, calculating () the distance therebetween for obtaining a corresponding second length; e) calculating a plurality of distance ratios (), each distance ratio corresponding to a selected pair of interest point matches and being based on a ratio of a first term and a second term or on a ratio of the second term and the first term, said first term corresponding to the distance between the first interest points of said pair of interest point matches and said second term corresponding to the distance between the second interest points of said pair of interest point matches; f) computing () a first representation of the statistical distribution of the plurality of calculated distance ratios; g) computing () a second representation of the statistical distribution of distance ratios obtained under the hypothesis that all the interest point matches in the image pair are outliers. The method further comprises generating () a first global representation of the statistical distribution of the plurality of calculated distance ratios computed for all the image pairs based on the first representations of all the image pairs; generating () a second global representation of the statistical distribution of distance ratios obtained under the hypothesis that all the interest point matches in all the image pairs are outliers based on the second representations of all the image pairs; comparing () said first global representation with said second global representation, and assessing () whether the first video shot contains a view of an object depicted in the second video shot based on said comparison.

Method and system for comparing video shots

A method for identifying keypoints in a digital image including a set of pixels. Each pixel has associated thereto a respective value of an image representative parameter. The method includes approximating a filtered image. The filtered image depends on a filtering parameter and includes for each pixel of the image a filtering function that depends on the filtering parameter to calculate a filtered value of the value of the representative parameter of the pixel. The approximating includes: a) generating a set of base filtered images; each base filtered image is the image filtered with a respective value of the filtering parameter; b) for each pixel of at least a subset of the set of pixels, approximating the filtering function by a respective approximation function based on the base filtered images; the approximation function is a function of the filtering parameter within a predefined range of the filtering parameter.

Keypoint identification

A method for generating a time scheduling of a contents stream distributed to plural terminals. The contents stream is segmented in consecutive segments, each segment defining a corresponding program. The method includes: receiving auxiliary metadata including a planned time scheduling of the content stream which has been set in advance and monitoring at least a portion of a segment of the contents stream to extract corresponding content features therefrom, generating a corresponding fingerprint based on the extracted content features and assessing at least one among an actual start time and an actual end time of such segment based on a comparison between the fingerprint and a reference fingerprint associated with such segment, and updating the planned time scheduling based on the assessed at least one among the actual start time and actual end time to generate a time scheduling synchronized with the contents stream.

Method and system for syncronizing electronic program guides

A method for processing an image, including: identifying a group of keypoints in the image; for each keypoint, calculating a corresponding descriptor array including plural array elements, each array element storing values taken by a corresponding color gradient histogram of a respective sub-region of the image in the neighborhood of the keypoint; for each keypoint, subdividing the descriptor array in at least two sub-arrays each including a respective number of elements of the descriptor array, and generating a compressed descriptor array including a corresponding compressed sub-array for each of the at least two sub-arrays, each compressed sub-array obtained by compressing the corresponding sub-array by vector quantization using a respective codebook; exploiting the compressed descriptor arrays of the keypoints for image analysis. For each keypoint of the group, the subdividing is based on correlation relationships among color gradient histograms with values stored in the elements of the descriptor array of each keypoint.

Method and system for image analysis based upon correlation relationships of sub-arrays of a descriptor array

A method for processing an image is proposed. The method comprises identifying a first group of keypoints in the image. For each keypoint of the first group, the method provides for identifying at least one corresponding keypoint local feature related to said each keypoint; for said at least one keypoint local feature, calculating a corresponding local feature relevance probability; calculating a keypoint relevance probability based on the local feature relevance probabilities of said at least one local feature. The method further comprises selecting keypoints, among the keypoints of the first group, having the highest keypoint relevance probabilities to form a second group of keypoints, and exploiting the keypoints of the second group for analyzing the image. The local feature relevance probability calculated for a local feature of a keypoint is obtained by comparing the value assumed by said local feature with a corresponding reference statistical distribution of values of said local feature.

Image analysis

A method for processing an image includes: identifying a group of keypoints in the image; for each keypoint of the group; a) calculating a corresponding descriptor array including a plurality of array elements, each array element storing values taken by a corresponding color gradient histogram of a respective sub-region of the image in the neighborhood of the keypoint; b) generating at least one compressed descriptor array by compressing at least one portion of the descriptor array by vector quantization using a codebook including a plurality of codewords.

Method and system for image analysis

A method of content recommendation, includes: generating a first digital mathematical representation of contents to associate the contents with a first plurality of words describing the contents; generating a second digital mathematical representation of text documents different from the contents to associate the documents with a second plurality of words; processing the first and second pluralities of words to determine a common plurality of words; processing the first and second digital mathematical representations to generate a common digital mathematical representation of the contents and the text documents based on the common plurality of words; and providing content recommendation by processing the common digital mathematical representation.

Method and system of content recommendation

A method for processing an image including: identifying a first group of keypoints in the image; for each keypoint of the first group, identifying at least one corresponding keypoint local feature related to the each keypoint; for the at least one keypoint local feature, calculating a corresponding local feature relevance probability; calculating a keypoint relevance probability based on the local feature relevance probabilities of the at least one local feature; selecting keypoints, among the keypoints of the first group, having the highest keypoint relevance probabilities to form a second group of keypoints, and exploiting the keypoints of the second group for analyzing the image. The local feature relevance probability calculated for a local feature of a keypoint is obtained by comparing the value assumed by the local feature with a corresponding reference statistical distribution of values of the local feature.

A method comparing first and second images, including: identifying and matching first and second keypoints in the first and second images; arranging a distribution of values of a calculated plurality of first distance ratios in a histogram; determining a number of correct keypoint matches, including: determining a matrix, each matrix element corresponding to a respective pair of keypoint matches with a value corresponding to a difference between a value of the histogram including a distance ratio of the respective pair of keypoint matches and an outlier probability density value weighted by a parameter; determining the parameter value such that the matrix dominant eigenvector is equal to a vector with a first value if the pair of keypoint match is correct and a second value if the pair of keypoint match is incorrect; determining the number of correct keypoint matches based on the dominant eigenvalue associated to the dominant eigenvector.

Method and system for comparing images

Method for determining the bounding box of a computerized graphic shape comprising one or more primitives described in vectorial mode, wherein the bounding box is calculated by applying analytical mathematical formulas to the vectorial description of the primitives constituting the shape without preliminarily performing a phase of drawing and displaying of the shape through resolution of the vectorial primitives that compose it.

Method for the calculation of the bounding box of vectorial graphic shapes

A method for comparing a first image with a second image. The method identifies first keypoints in the first image and second keypoints in the second image and associates each first keypoint with a corresponding second keypoint to form a corresponding keypoint match. For each pair of first keypoints, the method further calculates the distance therebetween for obtaining a corresponding first length. Similarly, for each pair of second keypoints, the method calculates the distance therebetween for obtaining a corresponding second length. The method further calculates a plurality of distance ratios; each distance ratio is based on a length ratio between a selected one between a first length and a second length and a corresponding selected one between a second length and a first length, respectively.

A method for estimating motion for the scalable video coding, includes the step of estimating the motion field of a sequence of photograms which can be represented with a plurality of space resolution levels including computing the motion field for the minimum resolution level and, until the maximum resolution level is reached, repeating the steps of: rising by one resolution level; extracting the photograms for such resolution level; and computing the motion field for such resolution level. The motion field is computed through an optical flow equation which contains, for every higher level than the minimum resolution level, a regularization factor between levels which points out the difference between the solution for the considered level and the solution for the immediately lower resolution level. A more or less high value of the regularization factor implies more or less relevant changes of the component at the considered resolution during the following process iterations.

Growing community of inventors

Turin, Italy

Gianluca Francini