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: A new tractable model solves labeling problems using regression tree fields, which represent non-parametric Gaussian conditional random fields. Regression tree fields are parameterized by non-parametric regression trees, allowing universal specification of interactions between image observations and variables. The new model uses regression trees corresponding to various factors to map dataset content (e.g., image content) to a set of parameters used to define the potential functions in the model. Some factors define relationships among multiple variable nodes. Further, the training of regression trees is scalable, both in the training set size and in the fact that the training can be parallelized. In one implementation, maximum pseudolikelihood learning provides for joint training of various aspects of the model, including feature test selection and ordering (i.e., the structure of the regression trees), parameters of each factor in the graph, and the scope of the interacting variable nodes used in the graph.

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: Methods of image segmentation using reduced foreground training data are described. In an embodiment, the foreground and background training data for use in segmentation of an image is determined by optimization of a modified energy function. The modified energy function is the energy function used in image segmentation with an additional term comprising a scalar value. The optimization is performed for different values of the scalar to produce multiple initial segmentations and one of these segmentations is selected based on pre-defined criteria. The training data is then used in segmenting the image. In other embodiments further methods are described: one places an ellipse inside the user-defined bounding box to define the background training data and another uses a comparison of properties of neighboring image elements, where one is outside the user-defined bounding box, to reduce the foreground training data.

Abstract: A method of up-sampling binary images for segmentation is described. In an embodiment, digital images are down-sampled before segmentation. The resulting initial binary segmentation, which has a lower resolution than the original image, is then up-sampled and smoothed to generate an interim non-binary solution which has a higher resolution than the initial binary segmentation. The final binary segmentation for the image is then computed from the interim non-binary solution based on a threshold. This method does not use the original image data in inferring the final binary segmentation solution from the initial binary segmentation. In an embodiment, the method may be applied to all images and in another embodiment, the method may be used for images which comprise a large number of pixels in total or in single dimension and smaller images may not be down-sampled before segmentation.

Abstract: A computing device is described herein that is configured to select a pixel pair including a foreground pixel of an image and a background pixel of the image from a global set of pixels based at least on spatial distances from an unknown pixel and color distances from the unknown pixel. The computing device is further configured to determine an opacity measure for the unknown pixel based at least on the selected pixel pair.

Abstract: Image completion using scene geometry is described, for example, to remove marks from digital photographs or complete regions which are blank due to editing. In an embodiment an image depicting, from a viewpoint, a scene of textured objects has regions to be completed. In an example, geometry of the scene is estimated from a depth map and the geometry used to warp the image so that at least some surfaces depicted in the image are fronto-parallel to the viewpoint. An image completion process is guided using distortion applied during the warping. For example, patches used to fill the regions are selected on the basis of distortion introduced by the warping. In examples where the scene comprises regions having only planar surfaces the warping process comprises rotating the image. Where the scene comprises non-planar surfaces, geodesic distances between image elements may be scaled to flatten the non-planar surfaces.

Abstract: Methods of image segmentation using reduced foreground training data are described. In an embodiment, the foreground and background training data for use in segmentation of an image is determined by optimization of a modified energy function. The modified energy function is the energy function used in image segmentation with an additional term comprising a scalar value. The optimization is performed for different values of the scalar to produce multiple initial segmentations and one of these segmentations is selected based on pre-defined criteria. The training data is then used in segmenting the image. In other embodiments further methods are described: one places an ellipse inside the user-defined bounding box to define the background training data and another uses a comparison of properties of neighboring image elements, where one is outside the user-defined bounding box, to reduce the foreground training data.

Abstract: A method of up-sampling binary images for segmentation is described. In an embodiment, digital images are down-sampled before segmentation. The resulting initial binary segmentation, which has a lower resolution than the original image, is then up-sampled and smoothed to generate an interim non-binary solution which has a higher resolution than the initial binary segmentation. The final binary segmentation for the image is then computed from the interim non-binary solution based on a threshold. This method does not use the original image data in inferring the final binary segmentation solution from the initial binary segmentation. In an embodiment, the method may be applied to all images and in another embodiment, the method may be used for images which comprise a large number of pixels in total or in single dimension and smaller images may not be down-sampled before segmentation.

Abstract: Methods of updating image segmentation following user input are described. In an embodiment, the properties used in computing the different portions of the image are updated as a result of one or more user inputs. Image elements which have been identified by a user input are given more weight when updating the properties than other image elements which have already been assigned to a particular portion of the image. In another embodiment, an updated segmentation is post-processed such that only regions which are connected to an appropriate user input are updated.

Abstract: Real-time segmentation of foreground from background layers in binocular video sequences may be provided by a segmentation process which may be based on one or more factors including likelihoods for stereo-matching, color, and optionally contrast, which may be fused to infer foreground and/or background layers accurately and efficiently. In one example, the stereo image may be segmented into foreground, background, and/or occluded regions using stereo disparities. The stereo-match likelihood may be fused with a contrast sensitive color model that is initialized or learned from training data. Segmentation may then be solved by an optimization algorithm such as dynamic programming or graph cut. In a second example, the stereo-match likelihood may be marginalized over foreground and background hypotheses, and fused with a contrast-sensitive color model that is initialized or learned from training data. Segmentation may then be solved by an optimization algorithm such as a binary graph cut.

Abstract: Real-time segmentation of foreground from background layers in binocular video sequences may be provided by a segmentation process which may be based on one or more factors including likelihoods for stereo-matching, color, and optionally contrast, which may be fused to infer foreground and/or background layers accurately and efficiently. In one example, the stereo image may be segmented into foreground, background, and/or occluded regions using stereo disparities. The stereo-match likelihood may be fused with a contrast sensitive color model that is initialized or learned from training data. Segmentation may then be solved by an optimization algorithm such as dynamic programming or graph cut. In a second example, the stereo-match likelihood may be marginalized over foreground and background hypotheses, and fused with a contrast-sensitive color model that is initialized or learned from training data. Segmentation may then be solved by an optimization algorithm such as a binary graph cut.

Abstract: Real-time segmentation of foreground from background layers in binocular video sequences may be provided by a segmentation process which may be based on one or more factors including likelihoods for stereo-matching, color, and optionally contrast, which may be fused to infer foreground and/or background layers accurately and efficiently. In one example, the stereo image may be segmented into foreground, background, and/or occluded regions using stereo disparities. The stereo-match likelihood may be fused with a contrast sensitive color model that is initialized or learned from training data. Segmentation may then be solved by an optimization algorithm such as dynamic programming or graph cut. In a second example, the stereo-match likelihood may be marginalized over foreground and background hypotheses, and fused with a contrast-sensitive color model that is initialized or learned from training data. Segmentation may then be solved by an optimization algorithm such as a binary graph cut.

Abstract: Techniques are disclosed to provide more efficient and improved extraction of a portion of a scene without requiring excessive user interaction. More particularly, the extraction may be achieved by using iterated graph cuts. In an implementation, a method includes segmenting an image into a foreground portion and a background portion (e.g., where an object or desired portion to be extracted is present in the foreground portion). The method determines the properties corresponding to the foreground and background portions of the image. Distributions may be utilized to model the foreground and background properties. The properties may be color in one implementation and the distributions may be a Gaussian Mixture Model in another implementation. The foreground and background properties are updated based on the portions. And, the foreground and background portions are updated based on the updated foreground and background properties.

Abstract: An output image formed from at least a portion of one or more input images may be automatically synthesized as a tapestry image. To determine which portion or region of each input image will be used in the image tapestry, the regions of each image may be labeled by one of a plurality of labels. The multi-class labeling problem of creating the tapestry may be resolved such that each region in the tapestry is constructed from one or more salient input image regions that are selected and placed such that neighboring blocks in the tapestry satisfy spatial compatibility. This solution may be formulated using a Markov Random Field and the resulting tapestry energy function may be optimized in any suitable manner. To optimize the tapestry energy function, an expansion move algorithm for energy functions may be generated to apply to non-metric hard and/or soft constraints.

Abstract: Techniques are disclosed to provide more efficient and improved border matting for extracted foreground images, e.g., without requiring excessive user interaction. Border matting techniques described herein generate relatively continuous transparency (or alpha values) along the boundary of the extracted object (e.g., limiting color bleeding and/or artifacts).