A processor implemented method of processing a facial expression image, the method includes controlling a camera to capture a first facial expression image and a second facial expression image, acquiring a first expression feature of the first facial expression image, acquiring a second expression feature of the second facial expression image, generating a new expression feature dependent on differences between the acquired first expression feature and the acquired second expression feature, and adjusting a target facial expression image based on the new expression feature.

An encoder is disclosed that uses hyperspectral data to produce a unified three-dimensional (“3D”) scan that incorporates depth for various points, surfaces, and features within a scene. The encoder may scan a particular point of the scene using frequencies from different electromagnetic spectrum bands, may determine spectral properties of the particular point based on returns measured across a first set of bands, may measure a distance of the particular point using frequencies of another band that does not interfere with the spectral properties at each of the first set of bands, and may encode the spectral properties and the distance of the particular point in a single hyperspectral dataset. The spectral signature encoded within the dataset may be used to classify the particular point or generate a point cloud or other visualization that accurately represents the spectral properties and distances of the scanned points.

A method is provided to using a machine learning model to predict lane geometry where incorrect or missing lane line geometry is detected. Methods may include: receiving a representation of lane line geometry for one or more roads of a road network; identifying an area within the representation including broken lane line geometry; generating a masked area of the area within the representation including the broken lane line geometry; processing the representation with the masked area through an inpainting model, where the inpainting model includes a generator network, where the representation is processed through the generator network which includes dilated convolution layers for inpainting of the masked area with corrected lane line geometry in a corrected representation; and updating a map database to include the corrected lane line geometry in place of the area including the broken lane line geometry based on the corrected representation.

A first map comprising local features and 3D locations of the local features is generated, the local features comprising visible features in a current image and a corresponding set of covisible features. A second map comprising prior features and 3D locations of the prior features may be determined, where each prior feature: was first imaged at a time prior to the first imaging of any of the local features, and lies within a threshold distance of at least one local feature. A first subset comprising previously imaged local features in the first map and a corresponding second subset of the prior features in the second map is determined by comparing the first and second maps, where each local feature in the first subset corresponds to a distinct prior feature in the second subset. A transformation mapping a subset of local features to a subset of prior features is determined.

Implementations are described herein for tracking objects with changing appearances across temporally-disparate images. In various implementations, a first probability distribution over a plurality of classes may be determined for a first biological object depicted in a first image captured at a first point in time. The classes may represent stages of growth of biological objects. Additional probability distribution(s) over the plurality of classes may be determined for candidate biological object(s) depicted in a second image captured at a second point in time subsequent to the first point in time. The candidate biological object(s) may potentially match the first biological object depicted in the first image. Based on a time interval between the first and second points in time, the first probability distribution may be compared to the probability distribution(s) of the candidate biological object(s) depicted in the second image to match one of the candidate biological object(s) depicted in the second image to the first biological object depicted in the first image.

One variation of a method for augmenting a digital procedure for production of pharmacological materials with automatic verification of objects includes: initializing a digital procedure containing a sequence of instructional blocks; and populating a instructional block with a instruction in a set of formats and a set of target objects. The method further includes, in response to an operator initiating an instance of the instructional block: accessing video feeds of the operator performing the instructional block; identifying a target object in a target frame from the video feeds; extracting visual features from the target frame; calculating an identification score for the target object based on the visual features; in response to the identification score falling below a confidence threshold, applying a fiducial to the target object; and linking the target object to the fiducial in the instructional block.

An object recognition method, apparatus, device and storage medium are provided. The method includes the following steps: a real-time image where a screen includes an object operation region is acquired; image reference points that match with at least two reference origins in the real-time image is determined according to the mapping relationship between the real-time image and a preset standard image, the reference origin is a point within the object operation region in the preset standard image; and an object within the object operation region is recognized according to a pixel coordinate of the preset standard image and a reference pixel coordinate of the image reference point in the real-time image to obtain a target recognition result of the object.

Techniques are described for determining whether a point cloud registration (e.g., alignment) between two sets of data points is valid. Such techniques can include determining voxelized representations of the sets of data points, and comparing characteristics of spatially aligned voxels within the voxelized representations. Characteristics of voxels to be compared can include classification labels of data points associated with voxels, including whether or not voxels correspond to free space. Point cloud registrations determined to be invalid can be given a weighting to be used in a subsequent high definition (HD) map building process. Generated maps can then be deployed for use in autonomous vehicles.

A three-dimensional information reconstraction device includes a corresponding point detector that detects a plurality of corresponding point pairs to which a first feature point included in a first image captured by a first image capturing device and a second feature point included in a second image captured by a second image capturing device correspond, and a three-dimensional coordinate deriver that, based on the plurality of corresponding point pairs, reconstructs three-dimensional coordinates to which the first feature point is inverse-projected.

Embodiments are provided to search for a dictionary image corresponding to a target image. The method includes detecting keypoints in a set of dictionary images. The set of dictionary images includes at least one dictionary image having a plurality of pixels. At least one random pair of pixels is selected among the detected keypoints of the dictionary image on the basis of candidate coordinates for pixels distributed around the detected keypoints of the dictionary image. A feature vector of each keypoint of the dictionary image is calculated, including calculating a difference in brightness between the selected pairs of pixels of the dictionary image. The calculated difference in brightness is an element of the feature vector. Keypoints of a target image are detected.