Amazon Technologies, Inc.

Nikhil Garg

Brad Bowman

Zohar Barzelay

Toma Belenzada

Michael Donoser

Sabine Sternig

Matthieu Guillaumin

Suren Kumar

Loris Bazzani

Bojan Pepik

Antonio D'Innocente

Systems and methods are provided for classifying images associated with an item, and generating an image set for that item which includes image classifications determined to be helpful for the item type of the item. To classify images, an image classification model is generated and trained using two phases. The first phase uses intermediate model with text and visual processing to teach the model to recognize patterns created by text without requiring OCR at inference. The second phase uses visual processing to refine the model for use at inference. To generate an image set, image classifications helpful to an item type are identified, items are associated with item types, images are obtained for an item, the images are classified using the image classification model, missing image classifications set out in the preferred image set are identified, and a request or requests is generated for the missing image classifications.

Image classification with modality dropout

Devices and techniques are generally described for retrieving three dimensional part models. Two-dimensional (2D) image data representing an object with at least a first part and a second part may be received. A first machine learning model may be used to generate a first shape embedding representing the first part and a second shape embedding representing the second part. A first 3D model stored in a non-transitory computer-readable memory that represents the first part may be determined based at least in part on the first shape embedding. A second 3D model stored in the non-transitory computer-readable memory that represents the second part may be determined based at least in part on the second shape embedding.

Three dimensional object part retrieval

Devices and techniques are generally described for determining localized visual similarity. In some examples, a selection of a first location of interest on a first image data depicting at least one article of clothing may be received. In some examples, a first machine learning model may generate a feature map representing the first image data. In some examples, a reduced feature map may be generated based at least in part on a mapping of the first location of interest to the feature map. In some examples, a second image depicting at least a second article of clothing may be determined based at least in part on the reduced feature map.

Localized visual similarity

Various embodiments of systems and methods allow a system to identify subsets of items by mixing and matching identified features in one or more other items. A system can identify features of items in an item database. The system can then calculate 'fingerprints' of these features which are vectors describing the characteristics of the features. The system can present a collection of items and a user can select an item of the collection. The user can then select positive features to include in a search and/or negative features to include in the search. The system can then do a search of the database for items that contain features similar to those positive features and do not contain features similar to those negative features. The user can select features through a variety of means.

Feature-based search

System and methods are provided for improved visual search systems that can use local features of images. The search system provides a user interface that enables a user to select areas or portions of an image to search on the selected features and the overall appearance of the image. The search system further provides customized user interfaces to exclude certain portions of images from the search algorithms. The search system can be used to detect potential intellectual property risks associated with items in an electronic catalog.

Feature based image detection

Techniques for prioritizing images associated with an item to display an appropriate image based on a query are described herein. For example, an attention score for an item attribute in an image of an item may be generated based at least in part on a model that uses one or more images of the item. The item attribute for the item associated with a query may be obtained. A plurality of items may be determined based at least in part on the item attribute being associated with the plurality of items where an individual item of the plurality of items includes a plurality of images of the item. The plurality of images of the individual item may be ranked based at least in part on corresponding attention scores associated with each image of the plurality of images.

Search result image selection techniques

A first graph is generated from a text data set, with graph nodes representing named entities in the data set and edges representing relationships between the named entities, and with edge weights indicating confidence levels. At least one cycle of the graph may be designated as inconsistent using a rule set. An edge may be selected for deletion from the first graph based on its presence in an inconsistent cycle, the cycle's weight, and/or on the edge weight. A representation of relationships indicated in the modified graph is provided programmatically.

Growing community of inventors

Berlin, Germany

Nikhil Garg