Abstract: Filtering sensor data is described, for example, where filters conditioned on a local appearance of the signal are predicted by a machine learning system, and used to filter the sensor data. In various examples the sensor data is a stream of noisy video image data and the filtering process denoises the video stream. In various examples the sensor data is a depth image and the filtering process refines the depth image which may then be used for gesture recognition or other purposes. In various examples the sensor data is one dimensional measurement data from an electric motor and the filtering process denoises the measurements. In examples the machine learning system comprises a random decision forest where trees of the forest store filters at their leaves. In examples, the random decision forest is trained using a training objective with a data dependent regularization term.

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: A method of sensing depth using an RGB camera. In an example method, a color image of a scene is received from an RGB camera. The color image is applied to a trained machine learning component which uses features of the 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 RGB camera. In various examples, the machine learning component comprises one or more entangled geodesic random decision forests.

Abstract: Surface segmentation from RGB and depth images is described. In one example, a computer receives an image of a scene. The image has pixels which each have an associated color value and an associated depth value representing a distance between from an image sensor to a surface in the scene. The computer uses the depth values to derive a set of three-dimensional planes present within the scene. A cost function is used to determine whether each pixel belongs to one of the planes, and the image elements are labeled accordingly. The cost function has terms dependent on the depth value of a pixel, and the color values of the pixels and at least one neighboring pixel. In various examples, the planes can be extended until they intersect to determine the extent of the scene, and pixels not belonging to a plane can be labeled as objects on the surfaces.

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: Crowdsourcing systems with machine learning are described, for example, to aggregate answers to a crowdsourced task in a manner achieving good accuracy even where observed data about past behavior of crowd members is sparse. In various examples a machine learning system jointly learns variables describing characteristics of both individual crowd workers and communities of the workers. In various examples, the machine learning system learns aggregated labels. In examples learnt variables describing characteristics of an individual crowd worker are related, by addition of noise, to learnt variables describing characteristics of a community of which the individual is a member. In examples the crowdsourcing system uses the learnt variables describing characteristics of individual workers and of communities of workers for any one or more of: active learning, targeted training of workers, targeted issuance of tasks, calculating and issuing rewards.

Abstract: A tractable model solves certain labeling problems by providing potential functions having arbitrary dependencies upon an observed dataset (e.g., image data). The model uses decision trees corresponding to various factors to map dataset content to a set of parameters used to define the potential functions in the model. Some factors define relationships among multiple variable nodes. When making label predictions on a new dataset, the leaf nodes of the decision tree determine the effective weightings for such potential functions. In this manner, decision trees define non-parametric dependencies and can represent rich, arbitrary functional relationships if sufficient training data is available. Decision trees training is scalable, both in the training set size and by parallelization. Maximum pseudolikelihood learning can provide for joint training of aspects of the model, including feature test selection and ordering, factor weights, and the scope of the interacting variable nodes used in the graph.

Abstract: Generating computer models of 3D objects is described. In one example, depth images of an object captured by a substantially static depth camera are used to generate the model, which is stored in a memory device in a three-dimensional volume. Portions of the depth image determined to relate to the background are removed to leave a foreground depth image. The position and orientation of the object in the foreground depth image is tracked by comparison to a preceding depth image, and the foreground depth image is integrated into the volume by using the position and orientation to determine where to add data derived from the foreground depth image into the volume. In examples, the object is hand-rotated by a user before the depth camera. Hands that occlude the object are integrated out of the model as they do not move in sync with the object due to re-gripping.

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: Learning image processing tasks from scene reconstructions is described where the tasks may include but are not limited to: image de-noising, image in-painting, optical flow detection, interest point detection. In various embodiments training data is generated from a 2 or higher dimensional reconstruction of a scene and from empirical images of the same scene. In an example a machine learning system learns at least one parameter of a function for performing the image processing task by using the training data. In an example, the machine learning system comprises a random decision forest. In an example, the scene reconstruction is obtained by moving an image capture apparatus in an environment where the image capture apparatus has an associated dense reconstruction and camera tracking system.

Abstract: An object detection system is disclosed herein. The object detection system allows detection of one or more objects of interest using a probabilistic model. The probabilistic model may include voting elements usable to determine which hypotheses for locations of objects are probabilistically valid. The object detection system may apply an optimization algorithm such as a simple greedy algorithm to find hypotheses that optimize or maximize a posterior probability or log-posterior of the probabilistic model or a hypothesis receiving a maximal probabilistic vote from the voting elements in a respective iteration of the algorithm. Locations of detected objects may then be ascertained based on the found hypotheses.

Abstract: An enhanced training sample set containing new synthesized training images that are artificially generated from an original training sample set is provided to satisfactorily increase the accuracy of an object recognition system. The original sample set is artificially augmented by introducing one or more variations to the original images with little to no human input. There are a large number of possible variations that can be introduced to the original images, such as varying the image's position, orientation, and/or appearance and varying an object's context, scale, and/or rotation. Because there are computational constraints on the amount of training samples that can be processed by object recognition systems, one or more variations that will lead to a satisfactory increase in the accuracy of the object recognition performance are identified and introduced to the original images.

Abstract: Using high-level attributes to guide image processing is described. In an embodiment high-level attributes of images of people such as height, torso orientation, body shape, gender are used to guide processing of the images for various tasks including but not limited to joint position detection, body part classification, medical image analysis and others. In various embodiments one or more random decision forests are trained using images where global variable values such as player height are known in addition to ground-truth data appropriate for the image processing task concerned. In some examples sequences of images are used where global variables are static or vary smoothly over the sequence. In some examples one or more trained random decision forests are used to find global variable values as well as output values for the task concerned such as joint positions or body part classes.

Abstract: Image labeling is described, for example, to recognize body organs in a medical image, to label body parts in a depth image of a game player, to label objects in a video of a scene. In various embodiments an automated classifier uses geodesic features of an image, and optionally other types of features, to semantically segment an image. For example, the geodesic features relate to a distance between image elements, the distance taking into account information about image content between the image elements. In some examples the automated classifier is an entangled random decision forest in which data accumulated at earlier tree levels is used to make decisions at later tree levels. In some examples the automated classifier has auto-context by comprising two or more random decision forests. In various examples parallel processing and look up procedures are used.

Abstract: Image labeling with global parameters is described. In an embodiment a pose estimation system executes automatic body part labeling. For example, the system may compute joint recognition or body part segmentation for a gaming application. In another example, the system may compute organ labels for a medical imaging application. In an example, at least one global parameter, for example body height is computed for each of the images to be labeled. In an example, the global parameter is used to modify an image labeling process. For example the global parameter may be used to modify the input image to a canonical scale. In another example, the global parameter may be used to adaptively modify previously stored parameters of the image labeling process. In an example, the previously stored parameters may be computed from a reduced set of training data.

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: A method and apparatus for processing video is disclosed. In an embodiment, image features of an object within a frame of video footage are identified and the movement of each of these features is tracked throughout the video footage to determine its trajectory (track). The tracks are analyzed, the maximum separation of the tracks is determined and used to determine a texture map, which is in turn interpolated to provide an unwrap mosaic for the object. The process may be iterated to provide an improved mosaic. Effects or artwork can be overlaid on this mosaic and the edited mosaic can be warped via the mapping, and combined with layers of the original footage. The effect or artwork may move with the object's surface.

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: Compressing motion fields is described. In one example video compression may comprise computing a motion field representing the difference between a first image and a second image, the motion field being used to make a prediction of the second image. In various examples of encoding a sequence of video data the first image, motion field and a residual representing the error in the prediction may be encoded rather than the full image sequence. In various examples the motion field may represented by its coefficients in a linear basis, for example a wavelet basis, and an optimization may be carried out to minimize the cost of encoding the motion field and maximize the quality of the reconstructed image while also minimizing the residual error. In various examples the optimized motion field may quantized to enable encoding.

Abstract: Mobile camera localization using depth maps is described for robotics, immersive gaming, augmented reality and other applications. In an embodiment a mobile depth camera is tracked in an environment at the same time as a 3D model of the environment is formed using the sensed depth data. In an embodiment, when camera tracking fails, this is detected and the camera is relocalized either by using previously gathered keyframes or in other ways. In an embodiment, loop closures are detected in which the mobile camera revisits a location, by comparing features of a current depth map with the 3D model in real time. In embodiments the detected loop closures are used to improve the consistency and accuracy of the 3D model of the environment.