Abstract: The claimed subject matter provides for systems and/or methods for identification of instances of an object of interest in 2D images by creating a database of 3D curve models of each desired instance and comparing an image of an object of interest against such 3D curve models of instances. The present application describes identifying and verifying the make and model of a car from a possibly single image—after the models have been populated with training data of test images of many makes and models of cars. In one embodiment, an identification system may be constructed by generating a 3D curve model by back-projecting edge points onto a visual hull reconstruction from silhouettes of an instance. The system and methods employ chamfer distance and orientation distance provides reasonable verification performance, as well as an appearance model for the taillights of the car to increase the robustness of the system.

Abstract: The claimed subject matter provides for systems and/or methods for identification of instances of an object of interest in 2D images by creating a database of 3D curve models of each desired instance and comparing an image of an object of interest against such 3D curve models of instances. The present application describes identifying and verifying the make and model of a car from a possibly single image—after the models have been populated with training data of test images of many makes and models of cars. In one embodiment, an identification system may be constructed by generating a 3D curve model by back-projecting edge points onto a visual hull reconstruction from silhouettes of an instance. The system and methods employ chamfer distance and orientation distance provides reasonable verification performance, as well as an appearance model for the taillights of the car to increase the robustness of the system.

Abstract: A database of images may be accessed. A feature set may be computed for each image, respectively. Each feature set includes feature integers quantized from interest points of a corresponding image. An initial set of clusters of the feature sets is found based on min hashes of the feature sets. Given the clusters of feature sets, descriptors for each of the clusters are computed, respectively, by selecting feature integers from among the feature sets in a cluster. The clusters are then refined by comparing at least some of the feature sets with at least some of the cluster descriptors, and based on such comparing adding some of the feature sets to clusters whose feature descriptors have similarity to the feature sets.

Abstract: A process for compressing and decompressing non-keyframes in sequential sets of contemporaneous video frames making up multiple video streams where the video frames in a set depict substantially the same scene from different viewpoints. Each set of contemporaneous video frames has a plurality frames designated as keyframes with the remaining being non-keyframes. In one embodiment, the non-keyframes are compressed using a multi-directional spatial prediction technique. In another embodiment, the non-keyframes of each set of contemporaneous video frames are compressed using a combined chaining and spatial prediction compression technique. The spatial prediction compression technique employed can be a single direction technique where just one reference frame, and so one chain, is used to predict each non-keyframe, or it can be a multi-directional technique where two or more reference frames, and so chains, are used to predict each non-keyframe.

Abstract: A database of images may be accessed. A feature set may be computed for each image, respectively. Each feature set includes feature integers quantized from interest points of a corresponding image. An initial set of clusters of the feature sets is found based on min hashes of the feature sets. Given the clusters of feature sets, descriptors for each of the clusters are computed, respectively, by selecting feature integers from among the feature sets in a cluster. The clusters are then refined by comparing at least some of the feature sets with at least some of the cluster descriptors, and based on such comparing adding some of the feature sets to clusters whose feature descriptors have similarity to the feature sets.

Abstract: An image-wide matting technique that involves modeling an image using a layered representation is presented. This representation includes a main pixel color layer, a secondary pixel color layer, an alpha layer and a noise layer. The four-layer representation is generated using a statistical model. Once generated, this representation can be used advantageously in a number of image editing operations.

Abstract: An image editing technique employing a layered representation of the image is presented. The image representation includes a main pixel color layer, a secondary pixel color layer, an alpha layer and a noise layer. Generally, the pixel values of one or more pixel locations of one or more of the layers of the image representation are manipulated to effect a change. Once changed, the layers are combined to produce a revised image.

Abstract: A system and process for compressing and decompressing multiple video streams depicting substantially the same dynamic scene from different viewpoints. Each frame in each contemporaneous set of video frames of the multiple streams is represented by at least a two layers—a main layer and a boundary layer. Compression of the main layers involves first designating one or more of these layers in each set of contemporaneous frames as keyframes. For each set of contemporaneous frames in time sequence order, the main layer of each keyframe is compressed using an inter-frame compression technique. In addition, the main layer of each non-keyframe within the frame set under consideration is compressed using a spatial prediction compression technique. Finally, the boundary layers of each frame in the current frame set are each compressed using an intra-frame compression technique. Decompression is generally the reverse of the compression process.

Abstract: A technique for estimating the optical flow between images of a scene and a segmentation of the images is presented. This involves first establishing an initial segmentation of the images and an initial optical flow estimate for each segment of each images and its neighboring image or images. A refined optical flow estimate is computed for each segment of each image from the initial segmentation of that image and the initial optical flow of the segments of that image. Next, the segmentation of each image is refined from the last-computed optical flow estimates for each segment of the image. This process can continue in an iterative manner by further refining the optical flow estimates for the images using their respective last-computed segmentation, followed by further refining the segmentation of each image using their respective last-computed optical flow estimates, until a prescribed number of iterations have been completed.

Abstract: A system and process for computing a 3D reconstruction of a scene from multiple images thereof, which is based on a color segmentation-based approach, is presented. First, each image is independently segmented. Second, an initial disparity space distribution (DSD) is computed for each segment, using the assumption that all pixels within a segment have the same disparity. Next, each segment's DSD is refined using neighboring segments and its projection into other images. The assumption that each segment has a single disparity is then relaxed during a disparity smoothing stage. The result is a disparity map for each image, which in turn can be used to compute a per pixel depth map if the reconstruction application calls for it.

Abstract: A system and process for computing a 3D reconstruction of a scene from multiple images thereof, which is based on a color segmentation-based approach, is presented. First, each image is independently segmented. Second, an initial disparity space distribution (DSD) is computed for each segment, using the assumption that all pixels within a segment have the same disparity. Next, each segment's DSD is refined using neighboring segments and its projection into other images. The assumption that each segment has a single disparity is then relaxed during a disparity smoothing stage. The result is a disparity map for each image, which in turn can be used to compute a per pixel depth map if the reconstruction application calls for it.

Abstract: A system and process for generating, and then rendering and displaying, an interactive viewpoint video in which a user can watch a dynamic scene while manipulating (freezing, slowing down, or reversing) time and changing the viewpoint at will. In general, the interactive viewpoint video is generated using a small number of cameras to capture multiple video streams. A multi-view 3D reconstruction and matting technique is employed to create a layered representation of the video frames that enables both efficient compression and interactive playback of the captured dynamic scene, while at the same time allowing for real-time rendering.

Abstract: A system and process for generating, and then rendering and displaying, an interactive viewpoint video in which a user can watch a dynamic scene while manipulating (freezing, slowing down, or reversing) time and changing the viewpoint at will. In general, the interactive viewpoint video is generated using a small number of cameras to capture multiple video streams. A multi-view 3D reconstruction and matting technique is employed to create a layered representation of the video frames that enables both efficient compression and interactive playback of the captured dynamic scene, while at the same time allowing for real-time rendering.

Abstract: A Bayesian two-color image demosaicer and method for processing a digital color image to demosaic the image in such a way as to reduce image artifacts. The method and system are an improvement on and an enhancement to previous demosaicing techniques. A preliminary demosaicing pass is performed on the image to assign each pixel a fully specified RGB triple color value. The final color value of pixel in the processed image is restricted to be a linear combination of two colors. Fully-specified RGB triple color values for each pixel in an image used to find two clusters represented favored two colors. The amount of contribution from these favored two colors on the final color value then is determined. The method and system also can process multiple images to improve the demosaicing results. When using multiple images, sampling can be performed at a finer resolution, known as super resolution.

Abstract: A feature symbol triplets object instance recognizer and method for recognizing specific objects in a query image. Generally, the recognizer and method find repeatable features in the image, and match the repeatable features between a query image and a set of training images. More specifically, the recognizer and method finds features in the query image and then groups all possible combinations of three features in to feature triplets. Small regions or “patches” in the query image, and an affine transformation is applied to the patches to identify any similarity between patches in a query image and training images. The affine transformation is computed using position of neighboring features in each feature triplet. Next, all similar patches are found, and then pairs of images are aligned to determine if the patches agree in the position of the object. If they do, then it is said that object is found and identified.

Abstract: An “Image Denoiser” provides a probabilistic process for denoising color images by segmenting an input image into regions, estimating statistics within each region, and then estimating a clean (or denoised) image using a probabilistic model of image formation. In one embodiment, estimated blur between each region is used to reduce artificial sharpening of region boundaries resulting from denoising the input image. In further embodiments, the estimated blur is used for additional purposes, including sharpening edges between one or more regions, and selectively blurring or sharpening one or more specific regions of the image (i.e., “selective focus”) while maintaining the original blurring between the various regions.

Abstract: A system and process for rendering and displaying an interactive viewpoint video is presented in which a user can watch a dynamic scene while manipulating (freezing, slowing down, or reversing) time and changing the viewpoint at will. The ability to interactively control viewpoint while watching a video is an exciting new application for image-based rendering. Because any intermediate view can be synthesized at any time, with the potential for space-time manipulation, this type of video has been dubbed interactive viewpoint video.

Abstract: A system and process for generating a two-layer, 3D representation of a digital or digitized image from the image and a pixel disparity map of the image is presented. The two layer representation includes a main layer having pixels exhibiting background colors and background disparities associated with correspondingly located pixels of depth discontinuity areas in the image, as well as pixels exhibiting colors and disparities associated with correspondingly located pixels of the image not found in these depth discontinuity areas. The other layer is a boundary layer made up of pixels exhibiting foreground colors, foreground disparities and alpha values associated with the correspondingly located pixels of the depth discontinuity areas. The depth discontinuity areas correspond to prescribed sized areas surrounding depth discontinuities found in the image using a disparity map thereof.

Abstract: A system and process for rendering and displaying an interactive viewpoint video is presented in which a user can watch a dynamic scene while manipulating (freezing, slowing down, or reversing) time and changing the viewpoint at will. The ability to interactively control viewpoint while watching a video is an exciting new application for image-based rendering. Because any intermediate view can be synthesized at any time, with the potential for space-time manipulation, this type of video has been dubbed interactive viewpoint video.

Abstract: A technique for estimating the optical flow between images of a scene and a segmentation of the images is presented. This involves first establishing an initial segmentation of the images and an initial optical flow estimate for each segment of each images and its neighboring image or images. A refined optical flow estimate is computed for each segment of each image from the initial segmentation of that image and the initial optical flow of the segments of that image. Next, the segmentation of each image is refined from the last-computed optical flow estimates for each segment of the image. This process can continue in an iterative manner by further refining the optical flow estimates for the images using their respective last-computed segmentation, followed by further refining the segmentation of each image using their respective last-computed optical flow estimates, until a prescribed number of iterations have been completed.