Abstract: Systems and methods are described for object detection within a digital image using a hierarchical softmax function. The method may include applying a first softmax function of a softmax hierarchy on a digital image based on a first set of object classes that are children of a root node of a class hierarchy, then apply a second (and subsequent) softmax functions to the digital image based on a second (and subsequent) set of object classes, where the second (and subsequent) object classes are children nodes of an object class from the first (or parent) object classes. The methods may then include generating an object recognition output using a convolutional neural network (CNN) based at least in part on applying the first and second (and subsequent) softmax functions. In some cases, the hierarchical softmax function is the loss function for the CNN.

Abstract: The present disclosure relates to an image merging system that automatically and seamlessly detects and merges missing people for a set of digital images into a composite group photo. For instance, the image merging system utilizes a number of models and operations to automatically analyze multiple digital images to identify a missing person from a base image, segment the missing person from the second image, and generate a composite group photo by merging the segmented image of the missing person into the base image. In this manner, the image merging system automatically creates merged group photos that appear natural and realistic.

Abstract: Certain embodiments involve a method for generating a search result. The method includes processing devices performing operations including receiving a query having a text input by a joint embedding model trained to generate an image result. Training the joint embedding model includes accessing a set of images and textual information. Training further includes encoding the images into image feature vectors based on spatial features. Further, training includes encoding the textual information into textual feature vectors based on semantic information. Training further includes generating a set of image-text pairs based on matches between image feature vectors and textual feature vectors. Further, training includes generating a visual grounding dataset based on spatial information. Training further includes generating a set of visual-semantic joint embeddings by grounding the image-text pairs with the visual grounding dataset.

Abstract: The present disclosure relates to an object selection system that accurately detects and automatically selects user-requested objects (e.g., query objects) in a digital image. For example, the object selection system builds and utilizes an object selection pipeline to determine which object detection neural network to utilize to detect a query object based on analyzing the object class of the query object. In addition, the object selection system can add, update, or replace portions of the object selection pipeline to improve overall accuracy and efficiency of automatic object selection within an image.

Abstract: The present disclosure relates to an object selection system that automatically detects and selects objects in a digital image based on natural language-based inputs. For instance, the object selection system can utilize natural language processing tools to detect objects and their corresponding relationships within natural language object selection queries. For example, the object selection system can determine alternative object terms for unrecognized objects in a natural language object selection query. As another example, the object selection system can determine multiple types of relationships between objects in a natural language object selection query and utilize different object relationship models to select the requested query object.

Abstract: The disclosure describes one or more implementations of a neural network architecture pruning system that automatically and progressively prunes neural networks. For instance, the neural network architecture pruning system can automatically reduce the size of an untrained or previously-trained neural network without reducing the accuracy of the neural network. For example, the neural network architecture pruning system jointly trains portions of a neural network while progressively pruning redundant subsets of the neural network at each training iteration. In many instances, the neural network architecture pruning system increases the accuracy of the neural network by progressively removing excess or redundant portions (e.g., channels or layers) of the neural network. Further, by removing portions of a neural network, the neural network architecture pruning system can increase the efficiency of the neural network.

Abstract: In order to provide monocular depth prediction, a trained neural network may be used. To train the neural network, edge detection on a digital image may be performed to determine at least one edge of the digital image, and then a first point and a second point of the digital image may be sampled, based on the at least one edge. A relative depth between the first point and the second point may be predicted, and the neural network may be trained to perform monocular depth prediction using a loss function that compares the predicted relative depth with a ground truth relative depth between the first point and the second point.

Abstract: Techniques are disclosed for parsing a source image, to identify segments of one or more objects within the source image. The parsing is carried out by an image parsing pipeline that includes three distinct stages comprising three respectively neural network models. The source image can include one or more objects. A first neural network model of the pipeline identifies a section of the source image that includes the object comprising a plurality of segments. A second neural network model of the pipeline generates, from the section of the source image, a mask image, where the mask image identifys one or more segments of the object. A third neural network model of the pipeline further refines the identification of the segments in the mask image, to generate a parsed image. The parsed image identifies the segments of the object, by assigning corresponding unique labels to pixels of different segments of the object.

Abstract: Keyword localization digital image search techniques are described. These techniques support an ability to indicate “where” a corresponding keyword is to be expressed with respect to a layout in a respective digital image resulting from a search query. The search query may also include an indication of a size of the keyword as expressed in the digital image, a number of instances of the keyword, and so forth. Additionally, the techniques and systems as described herein support real time search through use of keyword signatures.

Abstract: The present disclosure relates to shaping the architecture of a neural network. For example, the disclosed systems can provide a neural network shaping mechanism for at least one sampling layer of a neural network. The neural network shaping mechanism can include a learnable scaling factor between a sampling rate of the at least one sampling layer and an additional sampling function. The disclosed systems can learn the scaling factor based on a dataset while jointly learning the network weights of the neural network. Based on the learned scaling factor, the disclosed systems can shape the architecture of the neural network by modifying the sampling rate of the at least one sampling layer.

Abstract: Embodiments of the present invention provide systems, methods, and non-transitory computer storage media for parsing a given input referring expression into a parse structure and generating a semantic computation graph to identify semantic relationships among and between objects. At a high level, when embodiments of the preset invention receive a referring expression, a parse tree is created and mapped into a hierarchical subject, predicate, object graph structure that labeled noun objects in the referring expression, the attributes of the labeled noun objects, and predicate relationships (e.g., verb actions or spatial propositions) between the labeled objects. Embodiments of the present invention then transform the subject, predicate, object graph structure into a semantic computation graph that may be recursively traversed and interpreted to determine how noun objects, their attributes and modifiers, and interrelationships are provided to downstream image editing, searching, or caption indexing tasks.

Abstract: In implementations of generating descriptions of image relationships, a computing device implements a description system which receives a source digital image and a target digital image. The description system generates a source feature sequence from the source digital image and a target feature sequence from the target digital image. A visual relationship between the source digital image and the target digital image is determined by using cross-attention between the source feature sequence and the target feature sequence. The system generates a description of a visual transformation between the source digital image and the target digital image based on the visual relationship.

Abstract: In some embodiments, an application receives a request to execute a convolutional neural network model. The application determines the computational complexity requirement for the neural network based on the computing resource available on the device. The application further determines the architecture of the convolutional neural network model by determining the locations of down-sampling layers within the convolutional neural network model based on the computational complexity requirement. The application reconfigures the architecture of the convolutional neural network model by moving the down-sampling layers to the determined locations and executes the convolutional neural network model to generate output results.

Abstract: Object search techniques for digital images are described. In the techniques described herein, semantic features are extracted on a per-object basis form a digital image. This supports location of objects within digital images and is not limited to semantic features of an entirety of the digital image as involved in conventional image similarity search techniques. This may be combined with indications a location of the object globally with respect to the digital image through use of a global segmentation mask, use of a local segmentation mask to capture post and characteristics of the object itself, and so on.

Abstract: Methods and systems are provided for generating enhanced image. A neural network system is trained where the training includes training a first neural network that generates enhanced images conditioned on content of an image undergoing enhancement and training a second neural network that designates realism of the enhanced images generated by the first neural network. The neural network system is trained by determine loss and accordingly adjusting the appropriate neural network(s). The trained neural network system is used to generate an enhanced aesthetic image from a selected image where the output enhanced aesthetic image has increased aesthetics when compared to the selected image.

Abstract: Based on a received digital image and text, a neural network trained to identify candidate text placement areas within images may be used to generate a mask for the digital image that includes a candidate text placement area. A bounding box for the digital image may be defined for the text and based on the candidate text placement area, and the text may be superimposed onto the digital image within the bounding box.

Abstract: Techniques are disclosed for characterizing and defining the location of a copy space in an image. A methodology implementing the techniques according to an embodiment includes applying a regression convolutional neural network (CNN) to an image. The regression CNN is configured to predict properties of the copy space such as size and type (natural or manufactured). The prediction is conditioned on a determination of the presence of the copy space in the image. The method further includes applying a segmentation CNN to the image. The segmentation CNN is configured to generate one or more pixel-level masks to define the location of copy spaces in the image, whether natural or manufactured, or to define the location of a background region of the image. The segmentation CNN may include a first stage comprising convolutional layers and a second stage comprising pairs of boundary refinement layers and bilinear up-sampling layers.

Abstract: The present disclosure relates to an object selection system that automatically detects and selects objects in a digital image utilizing a large-scale object detector. For instance, in response to receiving a request to automatically select a query object with an unknown object class in a digital image, the object selection system can utilize a large-scale object detector to detect potential objects in the image, filter out one or more potential objects, and label the remaining potential objects in the image to detect the query object. In some implementations, the large-scale object detector utilizes a region proposal model, a concept mask model, and an auto tagging model to automatically detect objects in the digital image.

Abstract: Various embodiments describe frame selection based on training and using a neural network. In an example, the neural network is a convolutional neural network trained with training pairs. Each training pair includes two training frames from a frame collection. The loss function relies on the estimated quality difference between the two training frames. Further, the definition of the loss function varies based on the actual quality difference between these two frames. In a further example, the neural network is trained by incorporating facial heatmaps generated from the training frames and facial quality scores of faces detected in the training frames. In addition, the training involves using a feature mean that represents an average of the features of the training frames belonging to the same frame collection. Once the neural network is trained, a frame collection is input thereto and a frame is selected based on generated quality scores.

Abstract: Certain embodiments involve learning features of content items (e.g., images) based on web data and user behavior data. For example, a system determines latent factors from the content items based on data including a user's text query or keyword query for a content item and the user's interaction with the content items based on the query (e.g., a user's click on a content item resulting from a search using the text query). The system uses the latent factors to learn features of the content items. The system uses a previously learned feature of the content items for iterating the process of learning features of the content items to learn additional features of the content items, which improves the accuracy with which the system is used to learn other features of the content items.