Abstract: A Video Semantic Segmentation System (VSSS) is disclosed that performs accurate and fast semantic segmentation of videos using a set of temporally distributed neural networks. The VSSS receives as input a video signal comprising a contiguous sequence of temporally-related video frames. The VSSS extracts features from the video frames in the contiguous sequence and based upon the extracted features, selects, from a set of labels, a label to be associated with each pixel of each video frame in the video signal. In certain embodiments, a set of multiple neural networks are used to extract the features to be used for video segmentation and the extraction of features is distributed among the multiple neural networks in the set. A strong feature representation representing the entirety of the features is produced for each video frame in the sequence of video frames by aggregating the output features extracted by the multiple neural networks.

Abstract: A system and method for processing an image including the steps of receiving an input image having a plurality of pixels, wherein each of the plurality of pixels have one or more pixel characteristics; and processing the input image to generate an enhanced image by applying a pixel/image relationship to each of the plurality of pixels of the input image, wherein the pixel/image relationship is arranged to adjust the one or more pixel characteristics of each of the plurality of pixels of the input image.

Abstract: A non-transitory computer-readable recording medium has stored therein a program that causes a computer to execute a process comprising: acquiring a first feature from a machine learning model that estimates a first result of a target after a first period in response to an input of a first chronological state of the target in the first period, the first feature being a feature of the first chronological state; acquiring a second feature by inputting a second chronological state to the machine learning model, the second feature being a feature of the second chronological state in a second period including a period after the first result is determined; and generating, based on the first feature and the second feature, training data that takes the second chronological state as an explanatory variable and takes a second result as an objective variable, the second result being obtained by changing the determined first result.

Abstract: One example method for identifying useful segments in surgical videos includes accessing a video of a surgical procedure and user activities of a plurality of users who have watched the video of the surgical procedure. The user activities include operations performed during playback of the video. The method further includes dividing the video into multiple segments and determining a popularity score for each of the multiple segments based on the operations. Useful segments are identified from the segments based on the popularity scores. The method further includes generating metadata for the video of the surgical procedure to include an indication of the identified useful segments and associating the metadata with the video of the surgical procedure.

Abstract: A processor-implemented neural network method includes: determining, using a neural network, a feature vector based on a training image of a first class among a plurality of classes; determining, using the neural network, plural feature angles between the feature vector and class vectors of other classes among the plurality of classes; determining a margin based on a class angle between a first class vector of the first class and a second class vector of a second class, among the class vectors, and a feature angle between the feature vector and the first class vector; determining a loss value using a loss function including an angle with the margin applied to the feature angle and the plural feature angles; and training the neural network by updating, based on the loss value, either one or both of one or more parameters of the neural network and one or more of the class vectors.

Abstract: An information processing device includes a processor. This processor: obtains at least one of first sensor data output from a first sensor and used to determine an ambient environment of a device in which a third sensor is placed, and second sensor data used to determine an orientation of this device; determines a tilt of a plane in a sensing direction of the third sensor with respect to the orientation of the device based on the at least one of the first sensor data and the second sensor data; determines, in accordance with the tilt determined, a processing target area of third sensor data output from the third sensor and used for object detection processing in the sensing direction of the third sensor; and executes the object detection processing using the processing target area determined, of the third sensor data.

Abstract: Disclosed systems and methods can include capturing the sequence of images of a monitored region that includes a sub-region of interest, processing the sequence of images using heuristics and rules of an artificial intelligence model to identify the plurality of discrete parts that are associated with a type of a unified entity, and processing the sequence of images using the heuristics and the rules of the artificial intelligence model to virtually link together a group of the plurality of discrete parts that correspond to a specific embodiment of the unified entity that is present in the sub-region of interest, wherein the heuristics and the rules of the artificial intelligence model can be developed from a training process that includes the artificial intelligence model receiving sample images delineating exemplary discrete parts on exemplary embodiments of the unified entity.

Abstract: Systems and methods are provided for learning item types of items listed in an electronic repository, and for training a machine learning model to predict the item type of a given input item. For example, a machine learning model may be obtained or accessed that has been previously trained to classify an input item to a browse node. Vector representations of individual items assigned to different browse nodes may be obtained from an intermediate layer of the previously trained machine learning model, and a vector representation of individual browse nodes may then be generated based on the vector representations of individual items assigned to that browse node. A clustering algorithm may be applied to the browse node vector representations in order to identify clusters of similar browse nodes, where individual clusters may represent different unique item types.

Abstract: An input unit 132 receives an input of an image subject to a process. A processing unit subjects the image input to the input unit 132 to a process of a convolutional neural network in which a fully connected layer is excluded. The convolutional neural network in the processing unit 114 includes a convolutional layer and a pooling layer. An output unit outputs a result of the process in the processing unit 114. A filter of the convolutional layer in the convolutional neural network in the processing unit 114 is trained to learn the result of the process having a 1×1 spatial dimension.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for associating a new measurement of an object surrounding a vehicle with a maintained track. One of the methods includes receiving an object track for a particular object, receiving a new measurement characterizing a new object at a new time step, and determining whether the new object is the same as the particular object, comprising: generating a representation of the new object at the new and preceding time steps; generating a representation of the particular object at the new and preceding time steps; processing a first network input comprising the representations using a first neural network to generate an embedding of the first network input; and processing the embedding of the first network input using a second neural network to generate a predicted likelihood that the new object and the particular object are the same.

Abstract: A method for training a defect detector comprises: obtaining a first reference image of a first reference object, wherein the first reference object has a defect and the first reference image has a first label indicating the defect; training a reconstruction model according to a second reference image of a second reference object associated with the first reference object, wherein a defect level of the second reference object is in a tolerable range with an upper limit; obtaining a target image of a target object associated with the first reference object and the second reference object; generating a second label according to the target image, the reconstruction model and an error calculation procedure, wherein the second label comprises a defect of the target object; and training a defect detector by performing a machine learning algorithm according to the first reference image, the target image and the second label.

Abstract: Disclosed is a reconfigurable convolution engine for performing a convolution operation on an image. A data receiving module receives image data. A determination module determines a kernel size based on the image data, clock speed associated to the convolution engine and number of available on-chip resources. An allocation module allocates a plurality of instances based on the kernel size. Each instance of the plurality of instances further comprises a set of computing blocks operating concurrently. Each computing block is configured to perform convolution operation on the feature map of the image. An aggregation module aggregates the convolution output of each computing block for each instance of the plurality of instances to produce a convolution result for the image.

Abstract: A method and apparatus for latency control in an optical network. A management node such as an OLT in a PON sends a discovery message intending to prompt joining network nodes such as ONUs to send a response on a first wavelength during a quiet window established for this purpose. When a response is received, a secondary upstream-transmission wavelength is assigned to the ONU. When the ONU sends data upstream according to a schedule calculated by the ONT, which schedule may include transmission times using the assigned secondary wavelength. In this case, the assigned secondary wavelength will be scheduled using a relatively smaller or no quiet window. This scheduling may be determined in part by the service or services used by the ONU.

Abstract: An optical filter comprises an array of waveguides fabricated on an optical integrated circuit (PIC). The array comprises individual waveguides, each of which receive light inputs, e.g., individual taps of a multi-tap optical filter used in an interference cancellation circuit. Typically, the output(s) of the individual waveguides are located at an exit (edge) of the PIC. At least one second waveguide in the array is patterned on the PIC in a converged configuration such that the light transiting these waveguides co-propagates and interacts across given portions of the respective waveguides before exiting the waveguide array along a common facet, thereby generating or inhibiting one of intermodulation products, and harmonics. This structural configuration enables the generation of various modes of transmission at the PIC exit, enabling more efficient transfer of the energy, e.g., to an associated photodetector (PD) that provides conversion of the energy to the RF domain.

Abstract: A computer system is trained to understand audio-visual spatial correspondence using audio-visual clips having multi-channel audio. The computer system includes an audio subnetwork, video subnetwork, and pretext subnetwork. The audio subnetwork receives the two channels of audio from the audio-visual clips, and the video subnetwork receives the video frames from the audio-visual clips. In a subset of the audio-visual clips the audio-visual spatial relationship is misaligned, causing the audio-visual spatial cues for the audio and video to be incorrect. The audio subnetwork outputs an audio feature vector for each audio-visual clip, and the video subnetwork outputs a video feature vector for each audio-visual clip. The audio and video feature vectors for each audio-visual clip are merged and provided to the pretext subnetwork, which is configured to classify the merged vector as either having a misaligned audio-visual spatial relationship or not.

Abstract: This disclosure relates to an apparatus for object detection. The apparatus comprises a video camera, an object detector, and a controller. The video camera may be configured to generate a video stream of frames. The object detector may be configured to accept the video stream as input data and to perform object detection. The controller may be coupled to the video camera and the object detector. The controller may be configured to manage object detection in order to satisfy a performance metric and/or operate within an operational constraint.

Abstract: A method and a system are for calculating an output from a tomographic scrollable image stack, including a large number of generated sectional images of a tissue to be examined. In this context, a tomographic scrollable image stack is received. An output for a display is calculated. The output includes a primary image that is intended for representing the received image stack, and the output includes a secondary image with additional information. The secondary image is displayed overlaid on the primary image once an unhide signal or hide signal is received. In this context, a reference may be provided between the additional information of the secondary image and the slice of tissue for examination that is shown in the sectional image of the image stack.

Abstract: A MapReduce-based training framework exploits both data parallelism and model parallelism to scale training of complex models. Particular model architectures facilitate and benefit from use of such training framework. As one example, a machine-learned model can include a shared feature extraction portion configured to receive and process a data input to produce an intermediate feature representation and a plurality of prediction heads that are configured to receive and process the intermediate feature representation to respectively produce a plurality of predictions. For example, the data input can be a video and the plurality of predictions can be a plurality of classifications for content of the video (e.g., relative to a plurality of classes).

Abstract: The deep penetration of optical transmission from the very edges of the network with optical access networks to the very core with routing data within data centers before transmission has resulted in competing demands for increased functionality, reduced cost, enhanced manufacturability, and reduced footprint. At the same time monitoring and fault detection with prior art optical time domain reflectometry systems have not kept up to the demands of these networks and systems as they are expensive test equipment based solutions. It would be beneficial to provide embedded OTDR functionality within each transmitter, receiver or transceiver deployed within the network allowing every link to be monitored continuously. It would be further beneficial for such embedded OTDRs to meet the demands for lower cost, high volumes, and smaller footprints with enhanced manufacturability.

Abstract: A method for prediction of an indication of motion using input from an event-based camera includes receiving events captured by an event-based camera, wherein each of the events represents a location of a change in pixel intensity, a polarity of the change, and a time. The method further includes discretizing the events into time discretized event volumes, each of which contain events that occur within a specified time range. The method further includes providing the time discretized event volumes as input to an encoder-decoder neural network trained to predict an indication of motion using a loss function that measures quality of image deblurring; generating, using the neural network, a prediction of the indication of motion. The method further includes using the prediction of the indication of motion in a machine vision application.