Abstract: Various technologies described herein pertain to correction of an input depth image captured by a depth sensor. The input depth image can include pixels, and the pixels can have respective depth values in the input depth image. Moreover, per-pixel correction values for the pixels can be determined utilizing depth calibration data for a non-linear error model calibrated for the depth sensor. The per-pixel correction values can be determined based on portions of the depth calibration data respectively corresponding to the pixels and the depth values. The per-pixel correction values can be applied to the depth values to generate a corrected depth image. Further, the corrected depth image can be output.

Abstract: Various technologies described herein pertain to automatic in-situ calibration and registration of a depth sensor and a robotic arm, where the depth sensor and the robotic arm operate in a workspace. The robotic arm can include an end effector. A non-parametric technique for registration between the depth sensor and the robotic arm can be implemented. The registration technique can utilize a sparse sampling of the workspace (e.g., collected during calibration or recalibration). A point cloud can be formed over calibration points and interpolation can be performed within the point cloud to map coordinates in a sensor coordinate frame to coordinates in an arm coordinate frame. Such technique can automatically incorporate intrinsic sensor parameters into transformations between the depth sensor and the robotic arm. Accordingly, an explicit model of intrinsics or biases of the depth sensor need not be utilized.

Abstract: Data about a physical object in a real-world environment is automatically collected and labeled. A mechanical device is used to maneuver the object into different poses within a three-dimensional workspace in the real-world environment. While the object is in each different pose an image of the object is input from one or more sensors and data specifying the pose is input from the mechanical device. The image of the object input from each of the sensors for each different pose is labeled with the data specifying the pose and with information identifying the object. A database for the object that includes these labeled images can be generated. The labeled images can also be used to train a detector and classifier to detect and recognize the object when it is in an environment that is similar to the real-world environment.

Abstract: Data about a physical object in a real-world environment is automatically collected and labeled. A mechanical device is used to maneuver the object into different poses within a three-dimensional workspace in the real-world environment. While the object is in each different pose an image of the object is input from one or more sensors and data specifying the pose is input from the mechanical device. The image of the object input from each of the sensors for each different pose is labeled with the data specifying the pose and with information identifying the object. A database for the object that includes these labeled images can be generated. The labeled images can also be used to train a detector and classifier to detect and recognize the object when it is in an environment that is similar to the real-world environment.

Abstract: Various technologies described herein pertain to automatic in-situ calibration and registration of a depth sensor and a robotic arm, where the depth sensor and the robotic arm operate in a workspace. The robotic arm can include an end effector. A non-parametric technique for registration between the depth sensor and the robotic arm can be implemented. The registration technique can utilize a sparse sampling of the workspace (e.g., collected during calibration or recalibration). A point cloud can be formed over calibration points and interpolation can be performed within the point cloud to map coordinates in a sensor coordinate frame to coordinates in an arm coordinate frame. Such technique can automatically incorporate intrinsic sensor parameters into transformations between the depth sensor and the robotic arm. Accordingly, an explicit model of intrinsics or biases of the depth sensor need not be utilized.

Abstract: Various technologies described herein pertain to automatic in-situ calibration and registration of a depth sensor and a robotic arm, where the depth sensor and the robotic arm operate in a workspace. The robotic arm can include an end effector. A non-parametric technique for registration between the depth sensor and the robotic arm can be implemented. The registration technique can utilize a sparse sampling of the workspace (e.g., collected during calibration or recalibration). A point cloud can be formed over calibration points and interpolation can be performed within the point cloud to map coordinates in a sensor coordinate frame to coordinates in an arm coordinate frame. Such technique can automatically incorporate intrinsic sensor parameters into transformations between the depth sensor and the robotic arm. Accordingly, an explicit model of intrinsics or biases of the depth sensor need not be utilized.

Abstract: Various technologies described herein pertain to automatic in-situ calibration and registration of a depth sensor and a robotic arm, where the depth sensor and the robotic arm operate in a workspace. The robotic arm can include an end effector. A non-parametric technique for registration between the depth sensor and the robotic arm can be implemented. The registration technique can utilize a sparse sampling of the workspace (e.g., collected during calibration or recalibration). A point cloud can be formed over calibration points and interpolation can be performed within the point cloud to map coordinates in a sensor coordinate frame to coordinates in an arm coordinate frame. Such technique can automatically incorporate intrinsic sensor parameters into transformations between the depth sensor and the robotic arm. Accordingly, an explicit model of intrinsics or biases of the depth sensor need not be utilized.

Abstract: Various technologies described herein pertain to correction of an input depth image captured by a depth sensor. The input depth image can include pixels, and the pixels can have respective depth values in the input depth image. Moreover, per-pixel correction values for the pixels can be determined utilizing depth calibration data for a non-linear error model calibrated for the depth sensor. The per-pixel correction values can be determined based on portions of the depth calibration data respectively corresponding to the pixels and the depth values. The per-pixel correction values can be applied to the depth values to generate a corrected depth image. Further, the corrected depth image can be output.

Abstract: Systems, methods, and computer storage media having computer-executable instructions embodied thereon for rewriting queries and labeling word pairs. Queries are received and alternate words are identified for word pairs (i.e., query words and alternate words). Word pair links are presented to users and indicators are received based on actions taken by the users. Labels are assigned to the word pairs based on the indicators and communicated to a classifier.

Abstract: Systems, methods, and computer storage media having computer-executable instructions embodied thereon for rewriting queries and labeling word pairs. Queries are received and alternate words are identified for word pairs (i.e., query words and alternate words). Word pair links are presented to users and indicators are received based on actions taken by the users. Labels are assigned to the word pairs based on the indicators and communicated to a classifier.

Abstract: A system for organizing images includes an extraction component that extracts visual information (e.g., faces, scenes, etc.) from the images. The extracted visual information is provided to a comparison component which computes similarity confidence data between the extracted visual information. The similarity confidence data is an indication of the likelihood that items of extracted visual information are similar. The comparison component then generates a visual distribution of the extracted visual information based upon the similarity confidence data. The visual distribution can include groupings of the extracted visual information based on computed similarity confidence data. For example, the visual distribution can be a two-dimensional layout of faces organized based on the computed similarity confidence data—with faces in closer proximity faces computed to have a greater probability of representing the same person.

Abstract: Systems and methods are described for face recognition using discriminatively trained orthogonal rank one tensor projections. In an exemplary system, images are treated as tensors, rather than as conventional vectors of pixels. During runtime, the system designs visual features—embodied as tensor projections—that minimize intraclass differences between instances of the same face while maximizing interclass differences between the face and faces of different people. Tensor projections are pursued sequentially over a training set of images and take the form of a rank one tensor, i.e., the outer product of a set of vectors. An exemplary technique ensures that the tensor projections are orthogonal to one another, thereby increasing ability to generalize and discriminate image features over conventional techniques.

Abstract: Various technologies and techniques are disclosed for using user corrections to help improve handwriting recognition operations. The system tracks user corrections to recognition results. The system receives handwritten input from the user and performs a recognition operation to determine a top recognized word. The prior corrections made by the user are analyzed to calculate a ratio of times the user has corrected the top recognized word to a particular other word as opposed to correcting the particular other word to the top recognized word. If the ratio meets or exceeds a required minimum, then at least one secondary source is optionally analyzed to determine if the particular other word is used a certain number of times more frequently than the top recognized word in the secondary source. The system performs a swap of the top recognized word with the particular other word when the required criteria are met.

Abstract: Learning to reorder alternates based on a user's personalized vocabulary may be provided. An alternate list provided to a user for replacing words input by the user via a character recognition application may be reordered based on data previously viewed or input by the user (personal data). The alternate list may contain generic data, for example, words for possible substitution with one or more words input by the user. By using the user's personal data and statistical learning methodologies in conjunction with generic data in the alternate list, the alternate list can be reordered to present a top alternate that more closely reflect the user's vocabulary. Accordingly, the user is presented with a top alternate that is more likely to be used by the user to replace data incorrectly input.

Abstract: Various technologies and techniques are disclosed that identify possible incorrect recognition results. Handwritten input is received from a user. A recognition operation is performed on the handwritten input to produce an initial recognition result. A possible incorrect recognition is identified using the self-consistency process that identifies the possible incorrect recognition when the initial recognition result is not consistent with a normal writing style of the user. The self-consistency process performs a comparison of the initial recognition result with at least one sample previously provided by the user. If the comparison reveals that the initial recognition result is not consistent with the at least one sample, then the result is identified as possibly incorrect. A classifier confidence process can be alternatively or additionally used to identify a possible incorrect recognition result.

Abstract: Various technologies and techniques are disclosed for improving handwriting recognition using a neural network by allowing a user to provide samples. A recognition operation is performed on the user's handwritten input, and the user is not satisfied with the recognition result. The user selects an option to train the neural network on one or more characters to improve the recognition results. The user is prompted to specify samples for the certain character, word, or phrase, and the neural network is adjusted for the certain character, word, or phrase. Handwritten input is later received from the user. A recognition operation is performed on the handwritten input using the neural network that was adjusted for the certain character or characters.

Abstract: Learning to reorder alternates based on a user's personalized vocabulary may be provided. An alternate list provided to a user for replacing words input by the user via a character recognition application may be reordered based on data previously viewed or input by the user (personal data). The alternate list may contain generic data, for example, words for possible substitution with one or more words input by the user. By using the user's personal data and statistical learning methodologies in conjunction with generic data in the alternate list, the alternate list can be reordered to present a top alternate that more closely reflect the user's vocabulary. Accordingly, the user is presented with a top alternate that is more likely to be used by the user to replace data incorrectly input.

Abstract: Systems and methods are described for face recognition using discriminatively trained orthogonal rank one tensor projections. In an exemplary system, images are treated as tensors, rather than as conventional vectors of pixels. During runtime, the system designs visual features—embodied as tensor projections—that minimize intraclass differences between instances of the same face while maximizing interclass differences between the face and faces of different people. Tensor projections are pursued sequentially over a training set of images and take the form of a rank one tensor, i.e., the outer product of a set of vectors. An exemplary technique ensures that the tensor projections are orthogonal to one another, thereby increasing ability to generalize and discriminate image features over conventional techniques.

Abstract: A system for organizing images includes an extraction component that extracts visual information (e.g., faces, scenes, etc.) from the images. The extracted visual information is provided to a comparison component which computes similarity confidence data between the extracted visual information. The similarity confidence data is an indication of the likelihood that items of extracted visual information are similar. The comparison component then generates a visual distribution of the extracted visual information based upon the similarity confidence data. The visual distribution can include groupings of the extracted visual information based on computed similarity confidence data. For example, the visual distribution can be a two-dimensional layout of faces organized based on the computed similarity confidence data—with faces in closer proximity faces computed to have a greater probability of representing the same person.

Abstract: An improved system and method for personalizing recognition of an input method is provided. A trainable handwriting recognizer may be personalized by using ink written by the user and text authored by the user. The system includes a personalization service engine and a framework with interfaces for collecting, storing, and accessing user ink and authored information for training recognizers. The trainers of the system may include a text trainer for augmenting a recognizer's dictionary using text content and a shape trainer for tuning generic recognizer components using ink data supplied by a user. The trainers may load multiple trainer clients, each capable of training one or more specific recognizers. Furthermore, a framework is provided for supporting pluggable trainers. Any trainable recognizer may be dynamically personalized using the harvested information authored by the user and ink written by the user.