Xerox Corporation

Naver Corporation

Diane Larlus-Larrondo

Florent C Perronnin

Jon Almazan

Jean-Luc Meunier

Gabriela Csurka

Julien Perez

Jean-Marc Andreoli

Naila Murray

Albert Gordo Soldevila

Craig John Saunders

Adrien Gaidon

Jose Antonio Rodriguez-Serrano

Bojana Gajic

Mert Bulent Sariyildiz

Yangmuzi Zhang

Rui Hu

A method and system pre-trains a convolutional neural network for image recognition based upon masked language modeling by inputting, to the convolutional neural network, an image; outputting, from the convolutional neural network, a visual embedding tensor of visual embedding vectors; tokenizing a caption to create a list of tokens, at least one token having visual correspondence to the image received by the convolutional neural network; randomly selecting one of the tokens in the list of tokens to be masked, the selected token being taken as ground truth; computing, using a language model neural network, hidden representations of the tokens; using the hidden representation of the masked token, as a query vector, to pool the visual embedding vectors in the visual embedding tensor, attentively; predicting the masked token by mapping the pooled visual embedding vectors to the tokens; determining a prediction loss associated with the masked token; and back-propagating the prediction loss to the convolutional neural network to tune parameters thereof.

Method for training a convolutional neural network for image recognition using image-conditioned masked language modeling

A method for re-identification of a subject in an image by pre-training a convolutional neural network to recognize individuals within a closed set of possible identifications and further pre-training the convolutional neural network using classification loss; training the pre-trained convolutional neural network by sequentially processing a plurality of triplet of images, each triplet containing a query image degraded by adding random noise to a region of the query image, a positive image corresponding to an image of a same subject as in the query image, and a negative image corresponding to an image of a different subject as in the query image by (a) ranking the triplets by the triplet loss computed, (b) selecting a subset of triplets among the plurality of triplets, and (c) retraining the pre-trained convolutional neural network on each of the triplets of the subset of triplets.

Training and using a convolutional neural network for person re-identification

In a method for detecting an object in an input image, an input image vector representing the input image is generated by performing a regional maximum activations of convolutions (R-MAC) using a convolutional neural network (CNN) applied to the input image and using regions for the R-MAC defined by applying a region proposal network (RPN) to the output of the CNN applied to the input image. Likewise, a reference image vector representing a reference image depicting the object is generated by performing the R-MAC using the CNN applied to the reference image and using regions for the R MAC defined by applying the RPN to the output of the CNN applied to the reference image. A similarity metric between the input image vector and the reference image vector is computed, and the object is detected as present in the input image if the similarity metric satisfies a detection criterion.

Instance-level image retrieval with a region proposal network

An object detection method includes for each of a set of patches of an image, encoding features of the patch with a non-linear mapping function, and computing per-patch statistics based on the encoded features for approximating a window-level non-linear operation by a patch-level operation. Then, windows are extracted from the image, each window comprising a sub-set of the set of patches. Each of the windows is scored based on the computed patch statistics of the respective sub-set of patches. Objects, if any, can then be detected in the image, based on the window scores. The method and system allow the non-linear operations to be performed only at the patch level, reducing the computation time of the method, since there are generally many more windows than patches, while not impacting performance unduly, as compared to a system which performs non-linear operations at the window level.

Efficient object detection with patch-level window processing

A system and method for evaluating iconicity of an image are provided. In the method, at least one test image is received, each test image including an object in a selected class. Properties related to iconicity are computed for each test image. The properties may include one or more of: a) a direct measure of iconicity, which is computed with a direct iconicity prediction model which has been learned on a set of training images, each training image labeled with an iconicity score; b) one or more class-independent properties; and c) one or more class-dependent properties. A measure of iconicity of each of the test images is computed, based on the computed properties. By combining a set of complementary properties, an iconicity measure which shows good agreement with human evaluations of iconicity can be obtained.

System and method for predicting iconicity of an image

A system and method predict the translation quality of a translated input document. The method includes receiving an input document pair composed of a plurality of sentence pairs, each sentence pair including a source sentence in a source language and a machine translation of the source language sentence to a target language sentence. For each of the sentence pairs, a representation of the sentence pair is generated, based on a set of features extracted for the sentence pair. Using a generative model, a representation of the input document pair is generated, based on the sentence pair representations. A translation quality of the translated input document is computed, based on the representation of the input document pair.

Predicting the quality of automatic translation of an entire document

In an image classification method, a feature vector representing an input image is generated by unsupervised operations including extracting local descriptors from patches distributed over the input image, and a classification value for the input image is generated by applying a neural network (NN) to the feature vector. Extracting the feature vector may include encoding the local descriptors extracted from each patch using a generative model, such as Fisher vector encoding, aggregating the encoded local descriptors to form a vector, projecting the vector into a space of lower dimensionality, for example using Principal Component Analysis (PCA), and normalizing the feature vector of lower dimensionality to produce the feature vector representing the input image. A set of mid-level features representing the input image may be generated as the output of an intermediate layer of the NN.

Fisher vectors meet neural networks: a hybrid visual classification architecture

A computer implemented method for localization of an object, such as a license plate, in an input image includes generating a task-dependent representation of the input image based on relevance scores for the object to be localized. The relevance scores are output by a classifier for a plurality of locations in the input image, such as patches. The classifier is trained on patches extracted from training images and their respective relevance labels. One or more similar images are identified from a set of images, based on a comparison of the task-dependent representation of the input image and task-dependent representations of images in the set of images. A location of the object in the input image is identified based on object location annotations for the similar images.

Data driven localization using task-dependent representations

An image segmentation method includes generating a hierarchy of regions by unsupervised segmentation of an input image. Each region is described with a respective region feature vector representative of the region. Hierarchical structures are identified, each including a parent region and its respective child regions in the hierarchy. Each hierarchical structure is described with a respective hierarchical feature vector that is based on the region feature vectors of the respective parent and child regions. The hierarchical structures are classified according to a set of predefined classes with a hierarchical classifier component that is trained with hierarchical feature vectors of hierarchical structures of training images. The training images have semantic regions labeled according to the set of predefined classes. The input image is segmented into a plurality of semantic regions based on the classification of the hierarchical structures and optionally also on classification of the individual regions.

Growing community of inventors

La Tronche, France

Diane Larlus-Larrondo