A computer-implemented method for detecting landmark pairs in a pair of images is provided. The method includes receiving a pair of images, sampling the pair of images to generate reduced-resolution pairs of images, identifying features in the reduced-resolution pairs of images, matching the features in the image pairs, using the matched features in an increased resolution pair of images as guides for feature matching, and through iteratively guiding feature matching, identifying landmarks in the full-resolution pair of images.

Provided is an image processing device including: a processor comprising hardware, the processor configured to: smooth a brightness value of a cell image including a plurality of cell clusters each including a plurality of cells so as to generate a smoothed image in which a gap existing between the cells in each of the cell clusters is filled in; binarize the smoothed image into a background region and a non-background region of each cell cluster; and segment the non-background region of the binarized smoothed image into a region for each of the cell clusters.

System and techniques for positional analysis using computer vision sensor synchronization are described herein. A set of sensor data may be obtained for a participant of an activity. A video stream may be captured in response to detection of a start of the activity in the set of sensor data. The video stream may include images of the participant engaging in the activity. A key stage of the activity may be identified by evaluation of the sensor data. A key frame may be selected from the video stream using a timestamp of the sensor data used to identify the key stage of the activity. A skeletal map may be generated for the participant in the key frame using key points of the participant extracted from the key frame. Instructional data may be selected using the skeletal map. The instructional data may be displayed on a display device.

In some implementations, a device may detect edges in an image, and may identify, based on the edges, a rectangle that bounds a document in the image. The device may detect lines in the image, and may identify edge candidate lines by discarding one or more of the lines. The device may identify intersection points where lines, included in the edge candidate lines, intersect with one another. The device may identify corner candidate points by discarding one or more points included in the intersection points, and may identify a corner point included in the corner candidate points. The corner point may be a point, included in the corner candidate points, that is closest to one corner of the bounding rectangle. The device may perform perspective correction on the image of the document based on identifying the corner point.

The present disclosure generally provides a system and method for processing and identifying data in form. The system and method may distinguish between content data and background data in a form. In some aspects, the content data or background data may be removed, wherein the remaining data may be processed separately. Removal of the background data or the content data may allow for more effective or efficient character recognition of the data. In some embodiments, data may be processed on an element basis, wherein each element of the form may be labeled as background data, content data, noise, or combinations thereof. This system and method may significantly increase the ability to capture and extract relevant information from a form.

A controller may process a plurality of video frames to determine an apparent motion of each pixel of one or more pixels or each group of pixels of one or more groups of pixels of each video frame of the plurality of video frames. The controller may select one or more processed video frames, of the plurality of processed video frames, that correspond to a duration of time and may generate a composite video frame based on the one or more processed video frames. The controller may generate a video frame mask based on the composite video frame and may obtain additional video data that includes at least one additional video frame. The controller may cause the video frame mask to be applied to the at least one additional video frame and may cause the at least one additional video frame to be processed using an object detection technique.

Object re-identification refers to a process by which images that contain an object of interest are retrieved from a set of images captured using disparate cameras or in disparate environments. Object re-identification has many useful applications, particularly as it is applied to people (e.g. person tracking). Current re-identification processes rely on convolutional neural networks (CNNs) that learn re-identification for a particular object class from labeled training data specific to a certain domain (e.g. environment), but that do not apply well in other domains. The present disclosure provides cross-domain disentanglement of id-related and id-unrelated factors. In particular, the disentanglement is performed using a labeled image set and an unlabeled image set, respectively captured from different domains but for a same object class. The identification-related features may then be used to train a neural network to perform re-identification of objects in that object class from images captured from the second domain.

This application discloses a map element extraction method and apparatus, and a server. The map element extraction method includes obtaining a laser point cloud and an image of a target scene, the target scene including a map element; performing registration between the laser point cloud and the image to obtain a depth map of the image; performing image segmentation on the depth map of the image to obtain a segmented image of the map element in the depth map; and converting a two-dimensional location of the segmented image in the depth map to a three-dimensional location of the map element in the target scene according to a registration relationship between the laser point cloud and the image.

A method is provided for isolating and labeling discrete features in a spectral radiographic image recorded as a set of images in different energy channels. The disclosed method involves creating a profile for each of at least some pixels in the spectral radiographic image. The profiles are sequences of pixel values, in which each pixel value is a photon count or a similar radiographic exposure value indicative of the attenuation of a portion of the scanning beam in a respective energy channel. Iterative hierarchical clustering is used to cluster the pixels on the basis of their respective profiles. Labels are assigned to one or more of the resulting clusters. In implementations, each label can be associated with an inferred material composition or with an inference that the material composition is unknown.

An image processing apparatus includes a first generator configured to generate a first distance image corresponding to a first distance from an imager, by using an image captured by the imager; a second generator configured to generate a second distance image corresponding to a second distance that is further away from the imager than the first distance, by using an image captured by the imager; a reducer configured to reduce the first distance image; a first detector configured to detect a body positioned within the first distance, based on the first distance image reduced by the reducer; and a second detector configured to detect a body positioned within the second distance, based on the second distance image.