Abstract: Systems and methods for panoptic segmentation of a point cloud are provided. A point cloud is projected into a range image. Features are extracted from the range image and generating a feature map from the extracted features. The feature map is downsampled and the features are scaled during downsampling using local geometry. Features are extracted from the downsampled feature map. The point cloud is semantically segmented at least partially based on the features extracted. Instances in the point cloud are segmented at least partially based on the features extracted.

Abstract: Disclosed herein are a method and apparatus for compressing learning parameters for training of a deep-learning model and transmitting the compressed parameters in a distributed processing environment. Multiple electronic devices in the distributed processing system perform training of a neural network. By performing training, parameters are updated. The electronic device may share the updated parameter thereof with additional electronic devices. In order to efficiently share the parameter, the residual of the parameter is provided to the additional electronic devices. When the residual of the parameter is provided, the additional electronic devices update the parameter using the residual of the parameter.

Abstract: A laminated block correctness determining method determines whether a laminated block to be laminated is a correct laminated block when manufacturing a laminated iron core by laminating laminated blocks. Orientation identification portions are respectively disposed on opposite end surfaces of each of the laminated blocks in an axial direction. The laminated block correctness determining method includes acquiring, with an imaging device, a captured image of the laminated block by capturing one end surface of the laminated block in the axial direction. The laminated block correctness determining method further includes determining whether the orientation of the laminated block is correct by comparing the orientation identification portion in a registered image with the orientation identification portion in the captured image.

Abstract: An apparatus for lesion diagnosis and a method thereof are provided. The apparatus according to some example embodiments may perform acquiring a medical image of a subject, extracting a blood vessel region from the acquired medical image, measuring a distance between blood vessel bifurcation points in the extracted blood vessel region, and predicting a size of a lesion with respect to the measured distance. By doing this, the size of the lesion may be accurately predicted without intervention of human.

Abstract: A method includes: processing a captured original image based on an image signal processor (ISP) parameter, to obtain a target image; invoking, based on the target image, a pre-trained image quality assessment model to output and obtain a target assessment result, where the target assessment result indicates image quality of the assessed target image; and adjusting the ISP parameter based on the target assessment result.

Abstract: A system and method for unsupervised stereo matching with surface normal assistance for indoor applications. According to the disclosure, a deep neural network with a feature extraction module, a normal branch, and a disparity branch is disclosed. The extraction module and the normal branch are trained first in a supervised manner for surface normal prediction. The predicted surface normal is then incorporated into the disparity branch, which is trained later in an unsupervised manner for disparity estimation. The latter unsupervised learning approach can reduce our method's dependence on a large amount of ground truth data that is difficult to collect. Experimental results indicate that our proposed method can predict accurate surface normal at textureless regions. With the help of the surface normal, the predicted disparity at these challenging areas is more accurate, which leads to improved quality of stereo matching in indoor scenarios.

Abstract: An information processing apparatus includes one or more memories, and one or more processors that, when executing instructions stored in the one or more memories, function as the following units: an acquisition unit configured to acquire learning data including data and a label indicating a category of the data, a base holding unit configured to hold a base for generating a representative vector in the category, a learning unit configured to learn a parameter related to generation of the representative vector based on the acquired learning data, and a first generation unit configured to generate the representative vector based on the parameter and the base.

Abstract: Disclosed herein are training strategies for query-based object detectors, referred to herein as Query Recollection (QR). In one variation or QR, dense query recollection, every intermediate query is collected and independently forwarded to every downstream stage. In a second variation or QR, selective query recollection, intermediate queries are collected from the two nearest previous stages and forwarded to the next downstream stage. This eliminates the phenomena wherein intermediate stages of the decoder produce more accurate results than later stages of the decoder.

Abstract: A robotic system capable of being trained with a plurality of images that are synthetically augmented from an initial image data set includes a training system toward that end. An image augmentation system includes in one form a neural network trained to generate synthetic images using a generative adversarial network which includes the ability to synthesize images having various poses with adjustments to image parameters such as light and color among potential others. In another form the image augmentation system includes a set of images projected or transformed from its original pose to a number of different poses using an affine transform, and the ability to progress across an entire dimensional space of anticipated robot movements which produce various potential poses.

Abstract: In a depth estimation method, a depth estimation apparatus obtains a first color image, and inputs the first color image into a first depth estimation model to obtain a first intermediate depth image. The depth estimation apparatus then inputs the first color image and the first intermediate depth image into a second depth estimation model to obtain a first target depth image. The second depth estimation model is obtained through training based on a color image and a target depth image corresponding to the color image, and the first depth estimation model is obtained through training based on the color image and an intermediate depth image corresponding to the color image.

Abstract: Features extracted from one or more layers of a trained deep neural network (DNN) are used to detect out-of-distribution (OOD) data, such as anomalies. An OOD detection process includes transforming a feature output from a layer of the DNN from a relatively high-dimensional feature space to a lower-dimensional space, and then performing a reverse transformation back to the higher-dimensional feature space, resulting in a reconstructed feature. A feature reconstruction error is calculated based on a difference between the reconstructed feature and the original feature output from the DNN. The OOD detection process may further include calculating a score based on the feature reconstruction error and generating a visual representation of the feature reconstruction error.

Abstract: Methods and systems configured to embed data representing content output by at least one device for downstream content recognition or other downstream processes. For example, a device operates one or more media data embedding components configured to embed information regarding the output media, such as audio or image content. The embedding component is trained to embed information for a subset of known downstream processes with some known processes deliberately held out from the training. Among other benefits, this approach can help reduce over customization of the embedding component and allow more information to preserved by the component for purposes of downstream operations that may yet be configured.

Abstract: Techniques are described for self-supervised training of machine-learned models for detecting road features based on images of a driving environment. The model may be implemented as a neural network having two output heads: a first semantic segmentation head trained to output a likelihood of a particular road feature at various locations in the image, and a second geometric shape trained to output the parameters of a predicted road feature at the various locations. During training, the segmentation head of the model may be trained using labeled data associated with road features (e.g., driveways, lane geometry, etc.). The geometric shape head may be trained based on training data masked by the output of the segmentation head. In some examples, a Voronoi diagram may be generated to expand the labeled training data, and the segmentation output mask may be applied to the Voronoi diagram as the masked training data for the geometric shape head.

Abstract: A method for optimizing multi-frame processing model of a neural network includes: receiving a plurality of input frames by a processing engine that is configured to execute a multi frame processing model, the multi frame processing model including a plurality of convolution layers; selecting a pre-determined number of frames from the received plurality of frames for processing by the plurality of convolution layers; determining, as a sequence of frames, at least a preceding frame and a plurality of following frames amongst the selected pre-determined number of frames; removing the preceding frame by processing the sequence of frames using a plurality of filters in the multi frame processing model; and concatenating the plurality of following frames in an order, to the plurality of input frames for subsequent receiving by the multi frame processing model.

Abstract: A method and device for generating a composite group image from subgroup images is provided. Subgroup images, each having a common background, are accessed. The boundaries of a subgroup area within each of the subgroup images is determined. At least one horizontal and at least one vertical shift factor is determined using the determined boundaries. An arrangement for the subgroup images based on the at least one horizontal and the at least one vertical shift factor is generated. The composite group image is generated by blending the subgroup images arranged in the arrangement.

Abstract: A machine learning method includes: a first principal component analysis step configured to perform principal component analysis on learning data so as to thereby generate two or more principal components of the learning data; a first image data generation step configured to generate virtual image data by assigning, from among the two or more principal components, the first principal component to the X coordinate of an XY plane and the second principal component to the Y coordinate of the XY plane; and a learning step configured to generate a trained model by performing machine learning using the image data as input data.

Abstract: A computer implemented method of obtaining measurements of a person is disclosed herein. The method comprises obtaining a parametric deformable three-dimensional body model that can approximate the shape of any person, obtaining at least one image of the person, estimating one or more correspondences between the at least one image and the three-dimensional model, performing semantic image segmentation to segment the image of the person into their corresponding body parts, and iteratively adjusting at least one of (a) body pose and (b) shape parameters of the parametric deformable three-dimensional body model, to improve the fit of the three-dimensional model to at least one of: (i) the at least one image, (ii) the estimated one or more correspondences, and (iii) the segmented body parts. Measurements may then be extracted from the iteratively adjusted parametric deformable three-dimensional body model.

Abstract: A system and a method are disclosed for synthetic image generation. In some embodiments, the system includes one or more processors; and a memory storing instructions which, when executed by the one or more processors, cause performance of: receiving weather input data; receiving time input data; receiving pixel coordinates; using a light-source-modeling neural network, computing a light source model based on inputs of the weather input data and time input data; and using an image-generating system, generating an image based on the pixel coordinates and the light source model.

Abstract: A computer vision system, with at least one processor configured to: acquire, from a sports match video, moving image data of a first period and moving image data of a second period; by using a first machine learning model, generate, based on the moving image data of the first period, first estimation data and second estimation data for an estimation period; by using a second machine learning model, generate, based on the moving image data of the second period, the first estimation data and the second estimation data for the estimation period; and generate determination data based on the first estimation data and the second estimation data that are output from the first machine learning model and the first estimation data and the second estimation data that are output from the second machine learning model.

Abstract: An apparatus and method for determining driving risks of a driver using deep learning algorithms and a vehicle including the same are provided. The apparatus comprises a processor, a network interface, a memory, and a computer program loaded to the memory and executed by the processor, wherein the processor is configured to receive image data and CAN data obtained by a vehicle equipped with a lidar sensor or a camera sensor while the vehicle is driving, input the obtained image data and CAN data to a first deep learning algorithm trained through pre-stored image data to output image features related to driving risks of a driver driving the vehicle, output image features related to the driver's driving risk by the first deep learning algorithm, and capture a first image corresponding to the output image features and transmit the captured first image to a connect program.