Abstract: Embodiments relate to extracting features from images, such as by identifying keypoints and generating keypoint descriptors of the keypoints. An apparatus includes a pyramid image generator circuit, a keypoint descriptor generator circuit, and a pyramid image buffer. The pyramid image generator circuit generates an image pyramid from an input image. The keypoint descriptor generator circuit processes the pyramid images for keypoint descriptor generation. The pyramid image buffer stores different portions of the pyramid images generated by the pyramid image generator circuit at different times and provides the stored portions of the pyramid images to the keypoint descriptor generator circuit for keypoint descriptor generation. When first portions of the pyramid images are no longer needed for the keypoint descriptor generation, the first portions are removed from the pyramid image buffer to provide space for second portions of the pyramid images that are needed for the keypoint descriptor generation.

Abstract: A computer-implemented method for training a normalizing flow. The normalizing flow predicts a first density value based on a first input image. The first density value characterizes a likelihood of the first input image to occur. The first density value is predicted based on an intermediate output of a first convolutional layer of the normalizing flow. The intermediate output is determined based on a plurality of weights of the first convolutional layer. The method for training includes: determining a second input image; determining an output, wherein the output is determined by providing the second input image to the normalizing flow and providing an output of the normalizing flow as output; determining a second density value based on the output tensor and on the plurality of weights; determining a natural gradient of the plurality of weights with respect to the second density value; adapting the weights according to the natural gradient.

Abstract: A fully-automated computer-implemented system and method for generating a road network map from a remote sensing (RS) image in which the classification accuracy is satisfactory combines moving vehicle detection with spectral classification to overcome the limitations of each. Moving vehicle detections from an RS image are used as seeds to extract and characterize image-specific spectral roadway signatures from the same RS image. The RS image is then searched and the signatures matched against the scene to grow a road network map. The entire process can be performed using the radiance measurements of the scene without having to perform the complicated geometric and atmospheric conversions, thus improving computational efficiency, the accuracy of moving vehicle detection (location, speed, heading) and ultimately classification accuracy.

Abstract: Implementations are disclosed for selectively operating edge-based sensors and/or computational resources under circumstances dictated by observation of targeted plant trait(s) to generate targeted agricultural inferences. In various implementations, triage data may be acquired at a first level of detail from a sensor of an edge computing node carried through an agricultural field. The triage data may be locally processed at the edge using machine learning model(s) to detect targeted plant trait(s) exhibited by plant(s) in the field. Based on the detected plant trait(s), a region of interest (ROI) may be established in the field. Targeted inference data may be acquired at a second, greater level of detail from the sensor while the sensor is carried through the ROI. The targeted inference data may be locally processed at the edge using one or more of the machine learning models to make a targeted inference about plants within the ROI.

Abstract: Systems and methods for adapting physical activities and exercises based on facial analysis by image processing are disclosed. A method includes: identifying, by a computer device, a user; receiving, by the computer device, video data of a face region of the user while the user is engaged in exercise or physical activity; analyzing, by the computer device, the video data to determine a detected state of the user, wherein the analyzing the video data includes performing facial analysis using the video data; and providing, by the computer device, feedback to the user based on the analyzing the video data.

Abstract: Example solutions provide saliency for anchor-based object detection, and include: performing, with an object detector, a first object detection process on an image, wherein the first object detection process employs a plurality of anchor boxes; identifying an object detection result for the image; determining, from among the plurality of anchor boxes, a first anchor box associated with the object detection result; and while limiting the object detector to the first anchor box, generating, with the object detector, a saliency map for the image. In some examples, the saliency map is used for selecting further training data for the object detector. In some examples, the saliency map comprises a gradient-based saliency map, and is used for auditing or debugging the object detector.

Abstract: Provided is a system configured to obtain a set of images via a camera of the computing device, input the set of images into a neural network, and detect a target physical object with the neural network. The system may determine a contour of the target physical object and a first three-dimensional reconstruction of the target physical object. The system may generate a virtual representation and a virtual object based on attributes of the virtual representation, where a first attribute of the set of attributes includes the first three-dimensional reconstruction. The system may associate the virtual object with the virtual representation and displays the virtual object at pixel coordinates of a display that at least partially occlude at least part of the target physical object, where a position of the virtual object is computed based on the contour.

Abstract: A method of detecting an object in an image using a convolutional neural network (CNN) includes generating, based on the image, a plurality of reference feature maps and a corresponding feature pyramid including a plurality of final feature maps; obtaining a proposed region of interest (ROI); generating at least a first context ROI having an area larger than an area of the proposed ROI; assigning the proposed ROI and the first context ROI to a first and second final feature maps having different sizes; extracting, by performing ROI pooling, a first set of features from the first final feature map using the proposed ROI and a second set of features from the second final feature map using the first context ROI; and determining, based on the first and second sets of extracted features, at least one of a location of the object and a class of the object.

Abstract: Techniques for improving image processing related to item deliveries are described. In an example, a computer system receives an image showing a drop-off of an item, the item associated with a delivery to a delivery location. The computer system inputs the image to a first artificial intelligence (AI) model. The computer system receives first data comprising an indication of whether the drop-off is correct from the first AI model. The computer system causes a presentation of the indication at a device associated with the delivery of the item to the delivery location.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for tracking objects of interest using distance-based thresholding. One of the methods includes detecting an object depicted in an image captured by a camera, determining a predicted physical distance between the object and the camera, selecting, from a plurality of predetermined confidence thresholds, a confidence threshold for the predicted physical distance, each confidence threshold in the plurality of predetermined confidence thresholds for a different physical distance range, the confidence threshold having a physical distance range that includes the predicted physical distance, and determining, using the confidence threshold and a confidence score that indicates a likelihood that the object is an object of interest, that the object is likely an object of interest.

Abstract: A computer-implemented method of forecasting the semantic output of at least one frame, the method comprising the steps of receiving the input frames from a camera up to a predetermined time, processing via a down-sampling module of a neural network the plurality of input frames to receive a plurality of feature tensors, determining spatio-temporal correlations between the plurality of feature tensors, processing the plurality of feature tensors and the spatio-temporal correlations to receive at least one forecasted feature tensor, and processing via an up-sampling module of the neural network the at least one forecasted feature to receive at least one forecasted semantic output for a time larger than the predetermined time.

Abstract: Aspects of the subject technology relate to determining a holdup measurement based on a gamma spectrum through machine learning. A spectral image based on a gamma spectrum generated downhole in a wellbore can be accessed. A component of a holdup measurement for the wellbore can be classified into a specific quantized level through application of a machine learning classification model to the spectral image. A continuous value for the component of the holdup measurement for the wellbore can be quantified by applying a machine learning quantization model associated with the quantized level.

Abstract: A system for determining a location of a lighting fixture in a venue includes a camera and a controller. The camera captures images of the lighting fixture. The controller is operatively coupled to the camera. The controller designates a first position of the camera, receives a first image of the lighting fixture from the camera in the first position, determines a first distance vector between the first position and the lighting fixture based on the first image, designates a second position of the camera, receives a second image of the lighting fixture from the camera in the second position, determines a second distance vector between the second position and the lighting fixture based on the second image, determines a vector relationship location to be a location nearest both the first distance vector and the second distance vector, and designates the vector relationship location as the location of the lighting fixture.

Abstract: A system and a method for super-resolution image processing in remote sensing are disclosed. One or more sets of multi-temporal images with an input resolution and one or more first target images with a first output resolution are generated from one or more data sources. The first output resolution is higher than the input resolution. Each set of multi-temporal images is processed to improve an image match in the corresponding set of multi-temporal images. The one or more sets of multi-temporal images are associated with the one or more first target images to generate a training dataset. A deep learning model is trained using the training dataset. The deep learning model is provided for subsequent super-resolution image processing.

Abstract: An encoding apparatus is provided. The apparatus comprises an input unit operable to obtain a plurality of training images, said training images being for use in training a machine learning model. The apparatus also comprises a label unit operable to obtain a class label associated with the training images; and a key unit operable to obtain a secret key for use in encoding the training images. The apparatus further comprises an image noise generator operable to generate, based on the obtained secret key, noise for introducing into the training images. The image noise generator is configured to generate noise that correlates with the class label associated with the training images such that a machine learning model subsequently trained with the modified training images learns to associate the introduced noise with the class label for those images. A corresponding decoding apparatus is also provided.

Abstract: A method and an apparatus for generating an image are provided. The method includes: acquiring a screenshot of a webpage preloaded by a terminal as a source image; recognizing connection areas in the source image, and generating first circumscribed rectangular frames outside outlines of the connection areas; combining, in response to determining that a distance between the connection areas is smaller than a preset distance threshold, the connection areas, and generating a second circumscribed rectangular frame outside outlines of the combined connection areas; and generating, based on a nested relationship between the first circumscribed rectangular frames and the second circumscribed rectangular frames and pictures in the first circumscribed rectangular frames, a target image. The first circumscribed rectangular frames and the second circumscribed rectangular frame are respectively generated by recognizing and combining the connection areas in the source image.

Abstract: A method for removing fringe noise in an image includes: acquiring an original image; acquiring an original frequency spectrum of one-dimensional signal of the original image; determining a noise frequency band in the original frequency spectrum, and the noise frequency band is a frequency band including a central frequency of the fringe noise; denoising the noise frequency band to obtain a denoised frequency spectrum, wherein a denoising intensity used in the denoising is gradually increased from a start frequency position of the noise frequency band, and the denoising intensity is gradually reduced from the central frequency until a stop frequency position of the noise frequency band; and generating a denoised image according to the denoised frequency spectrum.

Abstract: A method for human-object interaction detection includes receiving an image. A set of features are extracted from multiple positions of the image. One or more human-object pairs may be predicted based on the extracted set of features. A human-object interaction may be determined based on a set of candidate interactions and the predicted human-object pairs.

Abstract: In a setting mode, a processor stores in a memory a reference image for pattern search, measurement location information indicating a plurality of measurement locations, and imaging setting information indicating at which magnification imaging is performed. In a measurement mode, the processor controls an imaging section, acquires a target image for pattern search, executes a pattern search on the target image using the reference image, specifies a plurality of imaging positions for imaging each of the plurality of measurement locations based on the measurement location information and a result of the pattern search, causes the imaging section to perform imaging at a magnification set for each measurement location, and measures a dimension of the measurement location.

Abstract: A target object detection model is provided. The target object detection model includes a YOLOv3-Tiny model. Through the target object detection model, low-level information in the YOLOv3-Tiny sub-model can be merged with high-level information therein, so as to fuse the low-level information and the high-level information. Since the low-level information can be further used, the comprehensiveness of target detection is effectively improved, and the detection effect of small targets is improved.