Various embodiments describe techniques for making inferences from graph-structured data using graph convolutional networks (GCNs). The GCNs use various pre-defined motifs to filter and select adjacent nodes for graph convolution at individual nodes, rather than merely using edge-defined immediate-neighbor adjacency for information integration at each node. In certain embodiments, the graph convolutional networks use attention mechanisms to select a motif from multiple motifs and select a step size for each respective node in a graph, in order to capture information from the most relevant neighborhood of the respective node.

A sketch-based image retrieval method, device and system, to improve accuracy of image searching from a scene sketch image. For example, the image retrieval method, device and system can be used to retrieve a target scene image from a collection of stored images in a storage (i.e., an image collection). The image retrieval method includes: segmenting the scene sketch image using an image segmentation module into semantic object-level instances, and fine-grained features are obtained for each object instance, generating an attribute graph which integrates the fine-grained features for each semantic object instance detected from the query scene sketch image, generating a feature graph by using a graph encoder module from the attribute graph, and computing a similarity or distance between the feature graphs of the query scene sketch image and the scene images in the image collection by a graph matching module and the most similar scene images are returned.

A method includes acquiring digital video data that portrays an interacting event, extracting image data, audio data, and semantic text data from the video data, analyzing the extracted data to identify a plurality of video features, and analyzing the plurality of video features to create a relationship graph. The interacting event comprises a plurality of interactions between plurality of individuals and the relationship graph comprises a plurality of nodes and a plurality of edges. Each node of the plurality of nodes represents an individual of the plurality of individuals, and each edge of the plurality of edges extends between two nodes of the plurality of nodes, and the plurality of edges represents the plurality of interactions. The method further comprises determining whether a first key feature is present in the relationship graph, wherein presence of the first key feature is predictive of a positive outcome of the interacting event.

A relative atlas may be used to lay out elements in a digital map used in the control of an autonomous vehicle. A vehicle pose for the autonomous vehicle within a geographical area may be determined, and the relative atlas may be accessed to identify elements in the geographical area and to determine relative poses between those elements. The elements may then be laid out within the digital map using the determined relative poses, e.g., for use in planning vehicle trajectories, for estimating the states of traffic controls, or for tracking and/or identifying dynamic objects, among other purposes.

Systems and methods are provided for augmenting digital analysis of lesions. An image of tissue having a glandular epithelial component is generated. The image represents a plurality of medium-scale epithelial components. For each of a plurality of cells within the image, a representative point is identified to provide a plurality of representative points for each of the plurality of medium-scale epithelial components. For each of a subset of the plurality of medium-scale epithelial components, a graph connecting the plurality of representative points is constructed. A plurality of classification features is extracted for each of the subset of medium-scale epithelial components from the graph constructed for the medium-scale epithelial component. A clinical parameter is assigned to each medium-scale epithelial component according to the extracted plurality of classification features.

Embodiments of an automated method of processing fingerprint images, identity information is extracted from prints typically classified as having “no identification value” because of sparse or missing minutiae by capturing ridge contour information. Bezier approximations of ridge curvature are used as Ridge Specific Markers. Control points arising from Bezier curves generate unique polygons that represent the actual curve in the fingerprint. The Bezier-based descriptors are then grouped together and compared to corresponding reference print Ridge Specific Marker data. The method makes it possible to fuse a plurality of individual latent print portions into a single descriptor of identity and use the resulting data for comparison and identification. Processing of poor quality reference prints according to the methods disclosed renders these prints useable for reference purposes.

Described herein are methods and systems for closed-form 3D model generation of non-rigid complex objects from scans with large holes. A computing device receives (i) a partial scan of a non-rigid complex object captured by a sensor coupled to the computing device; (ii) a partial 3D model corresponding to the object, and (iii) a whole 3D model corresponding to the object, wherein the partial 3D scan and the partial 3D model each includes one or more large holes. The device performs a rough match on the partial 3D model and changes the whole 3D model using the rough match to generate a deformed 3D model. The device refines the deformed 3D model using a deformation graph, reshapes the refined deformed 3D model to have greater detail, and adjusts the whole 3D model according to the reshaped 3D model to generate a closed-form 3D model that closes holes in the scan.

Methods, systems, and non-transitory computer readable storage media are disclosed for generating semantic scene graphs for digital images using an external knowledgebase for feature refinement. For example, the disclosed system can determine object proposals and subgraph proposals for a digital image to indicate candidate relationships between objects in the digital image. The disclosed system can then extract relationships from an external knowledgebase for refining features of the object proposals and the subgraph proposals. Additionally, the disclosed system can generate a semantic scene graph for the digital image based on the refined features of the object/subgraph proposals. Furthermore, the disclosed system can update/train a semantic scene graph generation network based on the generated semantic scene graph. The disclosed system can also reconstruct the image using object labels based on the refined features to further update/train the semantic scene graph generation network.

Methods, systems, and non-transitory computer readable storage media are disclosed for generating semantic scene graphs for digital images using an external knowledgebase for feature refinement. For example, the disclosed system can determine object proposals and subgraph proposals for a digital image to indicate candidate relationships between objects in the digital image. The disclosed system can then extract relationships from an external knowledgebase for refining features of the object proposals and the subgraph proposals. Additionally, the disclosed system can generate a semantic scene graph for the digital image based on the refined features of the object/subgraph proposals. Furthermore, the disclosed system can update/train a semantic scene graph generation network based on the generated semantic scene graph. The disclosed system can also reconstruct the image using object labels based on the refined features to further update/train the semantic scene graph generation network.

Techniques and systems are described for performing video segmentation using fully connected object proposals. For example, a number of object proposals for a video sequence are generated. A pruning step can be performed to retain high quality proposals that have sufficient discriminative power. A classifier can be used to provide a rough classification and subsampling of the data to reduce the size of the proposal space, while preserving a large pool of candidate proposals. A final labeling of the candidate proposals can then be determined, such as a foreground or background designation for each object proposal, by solving for a posteriori probability of a fully connected conditional random field, over which an energy function can be defined and minimized.