Abstract: Systems, apparatuses and methods may provide for technology that processes an inference workload in a first subset of layers of a neural network that prevents or inhibits data dependent branch operations, conducts an exit determination as to whether an output of the first subset of layers satisfies one or more exit criteria, and selectively bypasses processing of the output in a second subset of layers of the neural network based on the exit determination. The technology may also speculatively initiate the processing of the output in the second subset of layers while the exit determination is pending. Additionally, when the inference workloads include a plurality of batches, the technology may mask one or more of the plurality of batches from processing in the second subset of layers.

Abstract: Technical solutions for video event recognition relate to the fields of knowledge graphs, deep learning and computer vision. A video event graph is constructed, and each event in the video event graph includes: M argument roles of the event and respective arguments of the argument roles, with M being a positive integer greater than one. For a to-be-recognized video, respective arguments of the M argument roles of a to-be-recognized event corresponding to the video are acquired. According to the arguments acquired, an event is selected from the video event graph as a recognized event corresponding to the video.

Abstract: A method for optimizing an image classification model can include determining a first image classification model based on initial training data; in response to model optimization, determining a second image classification model based on the first image classification model and a noise data set; and obtaining a third image classification model by optimizing the second image classification model based on the initial training data, the third image classification model being configured to update the noise data set based on noise data generated within a predetermined time period and the noise data set.

Abstract: An information processing apparatus includes a processor configured to acquire a first recognition result and a first recognition probability on target data from a first recognizer, acquire a second recognition result and a second recognition probability on the target data from a second recognizer, execute checking of the first recognition result and the second recognition result, and execute first control in a case where the first recognition result and the second recognition result match each other as a result of the checking. The first control is control for executing either of first processing or second processing on the matched recognition result and outputting a processing result based on at least one of the first recognition probability or the second recognition probability. A human workload for the first processing is smaller than a human workload for the second processing.

Abstract: In some examples, a retroreflective article may include a retroreflective substrate, and an optical pattern embodied on the retroreflective substrate. The optical pattern may include a first optical sub-pattern and a second optical sub-pattern, wherein the optical pattern represents a set of information that is interpretable based on a combination of the first optical sub-pattern that is visible in a first light spectrum and the second optical sub-pattern that is visible in a second light spectrum. The first and second light spectra may be different.

Abstract: A method for reinforcement learning adopting a centering operation using a weight corresponding to a behavior probability is provided. The method includes steps of: a computing device instructing a reinforcement learning agent to (a) determine a k-th estimated parameter by referring to (i) a probability distribution for estimating the k-th behavior if k=1, and (ii) the probability distribution generated by selecting a (k?1)-th behavior if k>1; (b) select one of N behavior candidates as the k-th behavior by referring to (i) the k-th estimated parameter, and (ii) k-th contexts; (c) if k-th behavior probabilities are acquired, (i) generate k-th weighted expectation value of the k-th contexts, and (ii) generate k-th adjusted contexts by applying the centering operation to the k-th contexts; and (d) if a k-th reward is acquired, generate the probability distribution for estimating a (k+1)-th behavior using the k-th adjusted contexts, the k-th behavior probabilities, and the k-th reward.

Abstract: An object detection method and an associated electronic device are provided, wherein the object detection method includes: utilizing an image processing circuit to determine whether motion occurs in an image to generate a determination result; selectively utilizing a specific bounding box to identify a target object to generate an identification result according to the determination result, wherein the specific bounding box represents a location of the target object in a previous image; and selectively updating information of the specific bounding box according to the identification result.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for augmenting a neural network with additional operations.

Abstract: An electronic device is disclosed. The electronic device comprises a storage unit for storing a training model of a multi-dimensional long short-term memory (MDLSTM), and a processor for acquiring an image including at least one of handwritten text and printed text, identifying each text line region in the image through image processing, and recognizing text included in the each identified text line region, on the basis of the training model.

Abstract: A method and apparatus for performing facial registration includes selecting a plurality of target locations for registration. It is indicated that registration is taking place at each of the plurality of target locations, and it is indicated that registration is complete at the each of the plurality of target locations as each target location is registered. Once the registration of all target locations is complete, it is indicated to the system.

Abstract: A method comprising identifying a plurality of non-overlapping coarse domains of a region of interest; selecting a subset of the plurality of coarse domains based on a plurality of first diversity metrics determined for the plurality of coarse domains; identifying a plurality of non-overlapping fine domains of the region of interest, wherein each of the fine domains is a portion of one of the coarse domains of the selected subset of the plurality of coarse domains; selecting a subset of the plurality of fine domains based on a plurality of second diversity metrics determined for the plurality of coarse domains; and providing an indication of the selected subset of the plurality of fine domains.

Abstract: The embodiments of this disclosure disclose an image object detection method, device, electronic equipment, and computer-readable medium. A specific mode of carrying out the method includes: performing region segmentation on a target image to obtain at least one image region; performing feature extraction on each image region in the at least one image region to obtain at least one feature map; generating a semantic relation graph and a spatial distribution relation graph based on the at least one feature map and the at least one image region; generating an image region relation graph based on the semantic relation graph and spatial distribution relation graph; determining a target image region from the at least one image region based on the image region relation graph; displaying the target image region. This implementation mode achieves an improvement of user experience and a growth of network traffic.

Abstract: Methods and systems disclosed herein relate generally to systems and methods for predicting content based on vector data structures generated from image pixels. A content-prediction application accesses a color palette having two or more color-palette categories and selects a first color of the color palette. The content-prediction application generates a first vector based on a set of pixel values that represent the first color of the color palette. The content-prediction application determines a distance metric between the first vector and a second vector, in which the second vector is identified by applying a convolutional neural network model on an image depicting an item that includes a second color. In response to determining that the distance metric is less than a predetermined threshold, the content-prediction application selects the content item corresponding to the second vector.

Abstract: System and method of detecting infestation in agricultural products is disclosed. In one embodiment, an infestation detection system captures hyperspectral imaging data and depth imaging data for a plurality of points on an agricultural product upon directing a light source at the agricultural product. The system analyses the captured imaging data to derive morphological details as well as spectral signatures for complete 360° view of the plurality of points on the agricultural product. In an embodiment, the spectral signatures may be corrected for one or more pixels associated with the plurality of points in the agricultural product by integrating the hyperspectral imaging data with the depth imaging data. The system further classifies one or more regions of the agricultural product based on matching of spectral signatures for the one or more pixels with pre-defined spectral signatures stored for a plurality of agricultural products, and accordingly detect infestation.

Abstract: Systems and methods for operating a vehicle. The methods comprise, by a processor: obtaining a pair of stereo images captured by a stereo camera; processing the pair of stereo images to generate a disparity map comprising a plurality of pixels defined by intensity values; converting each intensity value to a 3D position in a map (each 3D position defining a location of a data point in a point cloud); performing a hierarchical decision tree classification to determine a classification for each data point in the point cloud (the classification being a foreground classification or a background classification); and using the classifications to facilitate autonomous control of the vehicle.

Abstract: An image processing apparatus acquires a plurality of captured images of characters captured in time series, each of the characters including a plurality of segments, recognize the characters captured for each of the plurality of captured images, and determine which one of the characters recognized from each of the plurality of captured images is to be output. The image processing apparatus determines, in accordance with a change aspect in time series of the characters recognized from each of the plurality of captured images, which one of the characters recognized from each of the plurality of captured images is to be output.

Abstract: The present disclosure provides an anomaly detection method and apparatus for multi-type data. According to the anomaly detection method for multi-type data, an adversarial learning network is trained, so that a generator in the adversarial learning network fits a distribution of a normal training sample and learns a potential mode of the normal training sample, to obtain an updated adversarial learning network, an anomaly evaluation function in the updated adversarial learning network is constructed according to a reconstruction error generated during training, and the updated adversarial learning network is constructed into an anomaly detection model, to perform anomaly detection on inputted detection data by the anomaly detection model, to obtain an anomaly detection result. A mode classifier is introduced to effectively resolve difficult anomaly detection when a distribution of detected data is similar to that of normal data, further improving the accuracy of anomaly detection.

Abstract: Embodiments of the present disclosure relate to a method, an apparatus, a device and a computer readable medium for generating VQA training data. The method comprises: obtaining a first group of training data of the VQA system, the first group of training data including a first question for an image in the VQA system and a first answer corresponding to the first question; determining a second question associated with the first question in term of semantic; and determining a second answer corresponding to the second question based on the first question and the first answer, to obtain a second group of training data for the image in the VQA system, the second group of training data including the second question and the second answer.

Abstract: A method of verifying a fault of an inspection unit, an inspection apparatus, and an inspection system are disclosed. The method according to the present disclosure includes: providing a verification reference body which is formed on a frame attached to an inspection system; placing the inspection unit on the verification reference body; obtaining image data of the verification reference body through the inspection unit; verifying a fault of the inspection unit by extracting a movement error and height error of the inspection unit from the image data; and generating a verification result indicating the fault of the inspection unit.

Abstract: A method for grasping texture-less metal parts based on BOLD image matching comprises: obtaining a real image and CAD template images by photographing, extracting a foreground part of the input part image, calculating a covariance matrix of a foreground image, establishing the direction of a temporary coordinate system, and setting directions of line segments to point to a first or second quadrant of the temporary coordinate system; constructing a descriptor of each line segment according to an angle relation between the line segment and k nearest line segments, and matching the descriptors of different line segments in the real image and the CAD template images to obtain line segment pairs; and recognizing a processed pose through a PNL algorithm to obtain a pose of a real texture-less metal part, and then inputting the pose of the real texture-less metal part to a mechanical arm to grasp the part.