Abstract: Tracking using sensor data is described, for example, where a plurality of machine learning predictors are used to predict a plurality of complementary, or diverse, parameter values of a process describing how the sensor data arises. In various examples a selector selects which of the predicted values are to be used, for example, to control a computing device. In some examples the tracked parameter values are pose of a moving camera or pose of an object moving in the field of view of a static camera; in some examples the tracked parameter values are of a 3D model of a hand or other articulated or deformable entity. The machine learning predictors have been trained in series, with training examples being reweighted after training an individual predictor, to favor training examples on which the set of predictors already trained performs poorly.

Abstract: Tracking hand or body pose from image data is described, for example, to control a game system, natural user interface or for augmented reality. In various examples a prediction engine takes a single frame of image data and predicts a distribution over a pose of a hand or body depicted in the image data. In examples, a stochastic optimizer has a pool of candidate poses of the hand or body which it iteratively refines, and samples from the predicted distribution are used to replace some candidate poses in the pool. In some examples a best candidate pose from the pool is selected as the current tracked pose and the selection processes uses a 3D model of the hand or body.

Abstract: In one or more implementations, a static geometry model is generated, from one or more images of a physical environment captured using a camera, using one or more static objects to model corresponding one or more objects in the physical environment. Interaction of a dynamic object with at least one of the static objects is identified by analyzing at least one image and a gesture is recognized from the identified interaction of the dynamic object with the at least one of the static objects to initiate an operation of the computing device.

Abstract: Memory facilitation using directed acyclic graphs is described, for example, where a plurality of directed acyclic graphs are trained for gesture recognition from human skeletal data, or to estimate human body joint positions from depth images for gesture detection. In various examples directed acyclic graphs are grown during training using a training objective which takes into account both connection patterns between nodes and split function parameter values. For example, a layer of child nodes is grown and connected to a parent layer of nodes using an initialization strategy. In examples, various local search processes are used to find good combinations of connection patterns and split function parameters.

Abstract: A method of controlling access to content such as web sites on the intranet or interne is described. For example, the method comprises receiving an address of content to be accessed and obtaining similarity of the address to previously labeled addresses of other content items. The similarity is based on co-occurrence of addresses of content items in records of browsing sessions from many consenting users. For example, a browsing session record comprises addresses of content items accessed by a user in a time period during which the user is actively accessing content. A co-occurrence of addresses of content items is the existence of the addresses in the same browsing session record. Access to the content is then controlled on the basis of the similarity.

Abstract: A method of sensing depth using an infrared camera. In an example method, an infrared image of a scene is received from an infrared camera. The infrared image is applied to a trained machine learning component which uses the intensity of image elements to assign all or some of the image elements a depth value which represents the distance between the surface depicted by the image element and the infrared camera. In various examples, the machine line component comprises one or more random decision forests.

Abstract: In one or more implementations, a static geometry model is generated, from one or more images of a physical environment captured using a camera, using one or more static objects to model corresponding one or more objects in the physical environment. Interaction of a dynamic object with at least one of the static objects is identified by analyzing at least one image and a gesture is recognized from the identified interaction of the dynamic object with the at least one of the static objects to initiate an operation of the computing device.

Abstract: Systems and methods for reducing interference between multiple infra-red depth cameras are described. In an embodiment, the system comprises multiple infra-red sources, each of which projects a structured light pattern into the environment. A controller is used to control the sources in order to reduce the interference caused by overlapping light patterns. Various methods are described including: cycling between the different sources, where the cycle used may be fixed or may change dynamically based on the scene detected using the cameras; setting the wavelength of each source so that overlapping patterns are at different wavelengths; moving source-camera pairs in independent motion patterns; and adjusting the shape of the projected light patterns to minimize overlap. These methods may also be combined in any way. In another embodiment, the system comprises a single source and a mirror system is used to cast the projected structured light pattern around the environment.

Abstract: Model fitting from raw time of flight image data is described, for example, to track position and orientation of a human hand or other entity. In various examples, raw image data depicting the entity is received from a time of flight camera. A 3D model of the entity is accessed and used to render, from the 3D model, simulations of raw time of flight image data depicting the entity in a specified pose/shape. The simulated raw image data and at least part of the received raw image data are compared and on the basis of the comparison, parameters of the entity are computed.

Abstract: Region of interest detection in raw time of flight images is described. For example, a computing device receives at least one raw image captured for a single frame by a time of flight camera. The raw image depicts one or more objects in an environment of the time of flight camera (such as human hands, bodies or any other objects). The raw image is input to a trained region detector and in response one or more regions of interest in the raw image are received. A received region of interest comprises image elements of the raw image which are predicted to depict at least part of one of the objects. A depth computation logic computes depth from the one or more regions of interest of the raw image.

Abstract: An inference engine is described for efficient machine learning. For example, an inference engine executes a plurality of ordered steps to carry out inference on the basis of observed data. For each step, a plurality of inputs to the step are received. A predictor predicts an output of the step and computes uncertainty of the prediction. Either the predicted output or a known output is selected on the basis of the uncertainty. If the known output is selected, the known output is computed, (for example, using a resource intensive, accurate process). The predictor is retrained using the known output and the plurality of inputs of the step as training data. For example, computing the prediction is fast and efficient as compared with computing the known output.

Abstract: Tracking hand or body pose from image data is described, for example, to control a game system, natural user interface or for augmented reality. In various examples a prediction engine takes a single frame of image data and predicts a distribution over a pose of a hand or body depicted in the image data. In examples, a stochastic optimizer has a pool of candidate poses of the hand or body which it iteratively refines, and samples from the predicted distribution are used to replace some candidate poses in the pool. In some examples a best candidate pose from the pool is selected as the current tracked pose and the selection processes uses a 3D model of the hand or body.

Abstract: Crowdsourcing using active learning is described, for example, to select pairs of tasks and groups of workers so that information gained about answers to the tasks in the pool is optimized. In various examples a machine learning system learns variables describing characteristics of communities of workers, characteristics of workers, task variables and uncertainty of these variables. In various examples, the machine learning system predicts task variables and uncertainty of the predicted task variables for possible combinations of communities of workers and tasks. In examples the predicted variables and uncertainty are used to calculate expected information gain of the possible combinations and to rank the possible combinations. In examples, the crowdsourcing system uses the expected information gain to allocate tasks to worker communities and observe the results; the results may then be used to update the machine learning system.

Abstract: Real-time camera tracking using depth maps is described. In an embodiment depth map frames are captured by a mobile depth camera at over 20 frames per second and used to dynamically update in real-time a set of registration parameters which specify how the mobile depth camera has moved. In examples the real-time camera tracking output is used for computer game applications and robotics. In an example, an iterative closest point process is used with projective data association and a point-to-plane error metric in order to compute the updated registration parameters. In an example, a graphics processing unit (GPU) implementation is used to optimize the error metric in real-time. In some embodiments, a dense 3D model of the mobile camera environment is used.

Abstract: Systems and methods for reducing interference between multiple infra-red depth cameras are described. In an embodiment, the system comprises multiple infra-red sources, each of which projects a structured light pattern into the environment. A controller is used to control the sources in order to reduce the interference caused by overlapping light patterns. Various methods are described including: cycling between the different sources, where the cycle used may be fixed or may change dynamically based on the scene detected using the cameras; setting the wavelength of each source so that overlapping patterns are at different wavelengths; moving source-camera pairs in independent motion patterns; and adjusting the shape of the projected light patterns to minimize overlap. These methods may also be combined in any way. In another embodiment, the system comprises a single source and a mirror system is used to cast the projected structured light pattern around the environment.

Abstract: Surface reconstruction contour completion embodiments are described which provide dense reconstruction of a scene from images captured from one or more viewpoints. Both a room layout and the full extent of partially occluded objects in a room can be inferred using a Contour Completion Random Field model to augment a reconstruction volume. The augmented reconstruction volume can then be used by any surface reconstruction pipeline to show previously occluded objects and surfaces.

Abstract: A “Layout Optimizer” provides various real-time iterative constraint-satisfaction methodologies that use constraint-based frameworks to generate optimized layouts that map or embed virtual objects into environments. The term environment refers to combinations of environmental characteristics, including, but not limited to, 2D or 3D scene geometry or layout, scene colors, patterns, and/or textures, scene illumination, scene heat sources, fixed or moving people, objects or fluids, etc., any of which may evolve or change over time. A set of parameters are specified or selected for each object. Further, the environmental characteristics are determined automatically or specified by users. Relationships between objects and/or the environment derived from constraints associated with objects and the environment are then used to iteratively determine optimized self-consistent and scene-consistent object layouts.

Abstract: Tracking using sensor data is described, for example, where a plurality of machine learning predictors are used to predict a plurality of complementary, or diverse, parameter values of a process describing how the sensor data arises. In various examples a selector selects which of the predicted values are to be used, for example, to control a computing device. In some examples the tracked parameter values are pose of a moving camera or pose of an object moving in the field of view of a static camera; in some examples the tracked parameter values are of a 3D model of a hand or other articulated or deformable entity. The machine learning predictors have been trained in series, with training examples being reweighted after training an individual predictor, to favour training examples on which the set of predictors already trained performs poorly.

Abstract: Surface reconstruction contour completion embodiments are described which provide dense reconstruction of a scene from images captured from one or more viewpoints. Both a room layout and the full extent of partially occluded objects in a room can be inferred using a Contour Completion Random Field model to augment a reconstruction volume. The augmented reconstruction volume can then be used by any surface reconstruction pipeline to show previously occluded objects and surfaces.

Abstract: Non-greedy machine learning for high accuracy is described, for example, where one or more random decision trees are trained for gesture recognition in order to control a computing-based device. In various examples, a random decision tree or directed acyclic graph (DAG) is grown using a greedy process and is then post-processed to recalculate, in a non-greedy process, leaf node parameters and split function parameters of internal nodes of the graph. In various examples the very large number of options to be assessed by the non-greedy process is reduced by using a constrained objective function. In examples the constrained objective function takes into account a binary code denoting decisions at split nodes of the tree or DAG. In examples, resulting trained decision trees are more compact and have improved generalization and accuracy.