Abstract: An information processing device includes a storage unit that stores learning information acquired by machine learning, an input unit that acquires identification information, and a processing unit that performs recognition processing using the learning information that is specified by the storage unit on a basis of the identification information. An information processing method is executed by a processor, the information processing method includes storing learning information acquired by machine learning, acquiring identification information, and performing recognition processing using the learning information that is specified from storage on a basis of the identification information.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing images using recurrent attention. One of the methods includes determining a location in the first image; extracting a glimpse from the first image using the location; generating a glimpse representation of the extracted glimpse; processing the glimpse representation using a recurrent neural network to update a current internal state of the recurrent neural network to generate a new internal state; processing the new internal state to select a location in a next image in the image sequence after the first image; and processing the new internal state to select an action from a predetermined set of possible actions.

Abstract: Technologies for using a shifted neural network include a compute device to determine a shift-based activation function of the shifted neural network. The shift-based activation function is a piecewise linear approximation of a transcendental activation function and is defined by a plurality of line segments such that a corresponding slope of each line segment is a power of two. The compute device further trains the shifted neural network based on shift-based input weights and the determined shift-based activation function.

Abstract: The present disclosure provides an image capture, curation, and editing system that includes a resource-efficient mobile image capture device that continuously captures images. In particular, the present disclosure provides low power frameworks for controlling image sensor mode in a mobile image capture device. On example low power frame work includes a scene analyzer that analyzes a scene depicted by a first image and, based at least in part on such analysis, causes an image sensor control signal to be provided to an image sensor to adjust at least one of the frame rate and the resolution of the image sensor.

Abstract: Disclosed is a system and method for receiving, at a user device, data from a near-field-communication tag on an object, initiating, based on the data, a browser on the user device, navigating, based on the data and via the browser, to a site and transmitting authorized payment data or other task to the site based on payment data retrieved from either the user device or a network entity. The data can be communicated from the browser to the site through an application programming interface. Any task can be performed as well such as opening a door, starting a car, or renting a parking space.

Abstract: A neuromorphic device including a convolution neural network is described. The convolution neural network may include an input layer having a plurality of input pixels, a plurality of kernel resistors, each of the kernel resistors corresponding to one of the plurality of input pixels, and an intermediate layer having a plurality of intermediate pixels electrically connected to the plurality of kernel resistors.

Abstract: The embodiments of the present disclosure provide an image processing method, and a processing device. The image processing method comprises: acquiring a first image including N components, where N is a positive integer greater than or equal to 1; and performing image conversion processing on the first image using a generative neural network, to output a first output image, wherein the generative neural network is trained using a Laplace transform function.

Abstract: Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency, such as accuracy of learning, accuracy of prediction, speed of learning, performance of learning, and energy efficiency of learning. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a respective compute element and a respective routing element. Each compute element has processing resources and memory resources. Each router enables communication via wavelets with at least nearest neighbors in a 2D mesh. Stochastic gradient descent, mini-batch gradient descent, and continuous propagation gradient descent are techniques usable to train weights of a neural network modeled by the processing elements. Reverse checkpoint is usable to reduce memory usage during the training.

Abstract: An electronic apparatus for compressing a language model is provided, the electronic apparatus including a storage configured to store a language model which includes an embedding matrix and a softmax matrix generated by a recurrent neural network (RNN) training based on basic data including a plurality of sentences, and a processor configured to convert the embedding matrix into a product of a first projection matrix and a shared matrix, the product of the first projection matrix and the shared matrix having a same size as a size of the embedding matrix, and to convert a transposed matrix of the softmax matrix into a product of a second projection matrix and the shared matrix, the product of the second projection matrix and the shared matrix having a same size as a size of the transposed matrix of the softmax matrix, and to update elements of the first projection matrix, the second projection matrix and the shared matrix by performing the RNN training with respect to the first projection matrix, the second p

Abstract: Embodiments of the present disclosure disclose a method and apparatus for facial recognition. A specific embodiment of the method includes: acquiring a to-be-recognized image; inputting the to-be-recognized image into a pre-trained first convolutional neural network to obtain complete facial feature information and partial facial feature information, the first convolutional neural network being used to extract a complete facial feature and a partial facial feature; and inputting the complete facial feature information and the partial facial feature information into a pre-trained second convolutional neural network to obtain a facial recognition result, the second convolutional neural network being used to represent a correlation between the facial recognition result, and the complete facial feature information and the partial facial feature information. This embodiment improves the accuracy of the recognition result in a situation where a face is partially covered.

Abstract: The described implementations enable a seller to sell items through multiple e-commerce channels without having to maintain independent merchant accounts at each channel. For example, a seller may sell items directly and through a management service. When a user request to purchase an item from the seller through the management service is received, the management service sends the purchase information to the seller so that the seller can complete the purchase as if the purchase were being made directly with the seller. Upon completion of the purchase, the seller provides a confirmation back to the management service and provides the item directly to the user.

Abstract: A first deep learning model is trained to classify general facial images. Cropped facial images are extracted from the general facial images. A second deep learning model is trained based on the cropped facial images. Face liveness detection is performed based on the trained first deep learning model and the trained second deep learning model.

Abstract: A semantic segmentation method and apparatus for improving an accuracy of a segmentation result are provided. The semantic segmentation method inputs a labeled image into a segmentation neural network to obtain segmentation information for the image, and back-propagates a segmentation loss for the segmentation information to update the segmentation neural network. The segmentation neural network is updated by further back-propagating an edge loss for the segmentation information.

Abstract: Systems and methods to receive one or more first images associated with a training set of images to train a machine learning model; provide the one or more first images as a first input to a first set of layers of computational units, wherein the first set of layers utilizes image filters; provide a first output of the first set of layers of computational units as a second input to a second layer of the computational units, wherein the second layer utilizes random parameter sets for computations; obtain distortion parameters from the second layer of the computational units; generate one or more second images comprising a representation of the one or more first images modified with the distortion parameters; obtain, as a third output, the one or more second images; and add the one or more second images to the training set of images to train the machine learning model.

Abstract: A mixed reality content providing apparatus is disclosed. The mixed reality content providing apparatus may recognize an OOI included in a 360-degree VR image to generate metadata of the OOI and may provide a user with mixed reality content where the metadata is overlaid on the 360-degree VR image.

Abstract: A mobile device can establish a communication with a separate device via a single function action such as bringing the devices near to each other. A method can include setting the mobile device to be in a state that enables a user to complete a purchase via the single function action with the mobile device, establishing, based on the user performing the single function action associated with the mobile device, a wireless link between the mobile device and the separate device, the wireless link providing communications associated with the purchase, receiving purchase data from the separate device via the wireless link, the purchase data being associated with the purchase, retrieving payment data from a memory of the mobile device and transmitting, from the mobile device to the separate device and via the wireless link, the payment data to make the purchase.

Abstract: A device can include a memory storing user payment data and another memory device storing instructions that cause the device to establish a communication between a separate device and the system based on a gesture associated with the system and via a wireless link between the system and the separate device, the communication being associated with a potential purchase, receive purchase data from the separate device via the wireless link, the purchase data being associated with the potential purchase, present, on the display, an instruction associated with the potential purchase, receive a single-interaction from the user of the system to confirm a payment for the potential purchase, the single-interaction comprising a security measure to prevent unauthorized purchases, retrieve, based on the single-interaction from the user, the user payment data from the memory and transmit the user payment data via the wireless link to the separate device to make a purchase.

Abstract: High resolution style transfer techniques and systems are described that overcome the challenges of transferring high resolution style features from one image to another image, and of the limited availability of training data to perform high resolution style transfer. In an example, a neural network is trained using high resolution style features which are extracted from a style image and are used in conjunction with an input image to apply the style features to the input image to generate a version of the input image transformed using the high resolution style features.

Abstract: Disclosed is a device including a processor, a computer-readable storage device and a software module stored on the computer-readable storage device, the software module configured with a software module application programming interface programmed that defines a protocol for communicating data between an application on the device and the software module. The module receives, from the application operating on the device, a request associated with a purchase from the application, wherein the request comprises information about the purchase, receives the authorized payment data and transmits, via the software module application programming interface and to the application, the authorized payment data.

Abstract: A memory control unit for handling data stored in a memory device includes a first interface to an interconnection with at least one memory bank; a second interface for communicating with a data requesting unit; and a memory quantization unit. The memory quantization unit is configured to: obtain, via the first interface, a first weight value from the at least one memory bank; quantize the first weight value to generate at least one quantized weight value having a shorter bit length than a bit length of the first weight value; and communicate the at least one quantized weight value to the data requesting unit via the second interface.

Abstract: The present disclosure provides a surveillance device that monitors an operation section of an automated transaction device, the surveillance device including: a learning holding section that learns and holds a reference scene data set in which a reference operation is divided into a sequence of action items; a feature extraction section that extracts actual target action data from actual scene data of the sequence of action items in an operation of a user, the actual scene data obtained from an imaging section that faces and images the operation section; and a detection section that associates actual target action data with a reference scene data set along the sequence of action items, compares for each of the action items, determines an anomaly level of the operation of the user, and outputs an anomalous occurrence signal according to the anomaly level.

Abstract: Methods and apparatus for improved deep learning for image acquisition are provided. An imaging system configuration apparatus includes a training learning device including a first processor to implement a first deep learning network (DLN) to learn a first set of imaging system configuration parameters based on a first set of inputs from a plurality of prior image acquisitions to configure at least one imaging system for image acquisition, the training learning device to receive and process feedback including operational data from the plurality of image acquisitions by the at least one imaging system. The example apparatus includes a deployed learning device including a second processor to implement a second DLN, the second DLN generated from the first DLN of the training learning device, the deployed learning device configured to provide a second imaging system configuration parameter to the imaging system in response to receiving a second input for image acquisition.

Abstract: An image recognition device includes: a hardware processor that: conducts machine learning, to perform a first process of calculating a plurality of region candidates for a region showing part of an object captured in an image, and a second process of determining a size of each of the region candidates in accordance with the object captured in the image; and determines the region from among the region candidates, using a predetermined criterion.

Abstract: Provided are systems, methods, and computer-readable medium for operating a neural network. In various implementations, the neural network can receive an input image that includes an object to be identified. The neural network can generate a plurality of initial feature maps using a convolution layers, wherein a first initial feature maps is generated using the input image. The neural network can generate an up-sampled feature map using a de-convolution layer that takes an initial feature map as input, where the up-sampled feature map has a same resolution as the previous initial feature map. The neural network can combine the up-sampled feature map and the previous initial feature map, and use the combined feature map to more accurate identify the object.

Abstract: Disclosed is an approach for enabling a user to choose from multiple payment options using a browser API. The method includes transmitting, to a browser and via a browser payment request application programming interface, a payment request having data associated with a purchase of a product from the site for a user and presenting a choice between a first payment method and a second payment method for purchasing the product. The method includes receiving a selection of a payment method from the user of one of the first payment method and the second payment method to yield a selected payment method and, based on the selected payment method and in response to the payment request, receiving, from the browser and via the browser payment request application programming interface, data associated with the selected payment method.

Abstract: A method is presented for generating a plurality of physical design layout patterns. The method includes selecting one or more physical design layouts for neural network training, converting the plurality of physical design layout patterns into coordinate arrays, a coordinate array of the coordinate arrays including via center coordinates of vias in a physical design layout pattern of the plurality of physical design layout patterns, training, by employing the coordinate arrays, a variational autoencoder (VAE), and generating one or more new synthetic coordinate arrays by employing the trained VAE, a synthetic coordinate array of the one or more new synthetic coordinate arrays including via center coordinates of vias for a new physical design layout pattern.

Abstract: A system estimates quality of a digital image by accessing a corpus of digital images of one or more subjects, such as a facet of a property. The system will receive, for at least a subset of the corpus, an indicator that one or more patches of each image in the subset is out of focus. The system will train a classifier by obtaining a feature representation of each pixel in each image, along with a focus value that represents an extent to which each pixel in the image is in focus or out of focus. The system will use the classifier to analyze pixels of a new digital image and assess whether each analyzed pixel in the new digital image is in focus or out of focus. The system may use the image to assess whether an incident occurred, such as storm-related damage to the property.

Abstract: Methods and systems are provided for using a single image of an indoor scene to estimate illumination of an environment that includes the portion captured in the image. A neural network system may be trained to estimate illumination by generating recovery light masks indicating a probability of each pixel within the larger environment being a light source. Additionally, low-frequency RGB images may be generated that indicating low-frequency information for the environment. The neural network system may be trained using training input images that are extracted from known panoramic images. Once trained, the neural network system infers plausible illumination information from a single image to realistically illumination images and objects being manipulated in graphics applications, such as with image compositing, modeling, and reconstruction.

Abstract: The present disclosure relates to a font recognition system that employs a multi-task learning framework and adversarial training to improve font classification and remove negative side effects caused by intra-class variances of glyph content. For example, in one or more embodiments, the font recognition system adversarial trains a font recognition neural network by minimizing font classification loss while at the same time maximizing glyph classification loss. By employing an adversarially trained font classification neural network, the font recognition system can improve overall font recognition by removing the negative side effects from diverse glyph content.

Abstract: A method for evaluating a reliability of labeling training images to be used for learning a deep learning network is provided. The method includes steps of: a reliability-evaluating device instructing a similar-image selection network to select validation image candidates with their own true labels having shooting environments similar to those of acquired original images, which are unlabeled images, and instructing an auto-labeling network to auto-label the validation image candidates with their own true labels and the original images; and (i) evaluating a reliability of the auto-labeling network by referring to true labels and auto labels of easy-validation images, and (ii) evaluating a reliability of a manual-labeling device by referring to true labels and manual labels of difficult-validation images. This method can be used to recognize surroundings by applying a bag-of-words model, to optimize sampling processes for selecting a valid image among similar images, and to reduce annotation costs.

Abstract: A method, computer readable medium, and system are disclosed for training a neural network. The method includes the steps of selecting an input sample from a set of training data that includes input samples and noisy target samples, where the input samples and the noisy target samples each correspond to a latent, clean target sample. The input sample is processed by a neural network model to produce an output and a noisy target sample is selected from the set of training data, where the noisy target samples have a distribution relative to the latent, clean target sample. The method also includes adjusting parameter values of the neural network model to reduce differences between the output and the noisy target sample.

Abstract: A classification engine stores a plurality of neural networks in memory, where each neural network is trained to recognize a set of confused graphemes from one or more sets of confused graphemes identified in recognition data for a plurality of document images. The classification engine receives an input grapheme image associated with a document image comprising a plurality of graphemes, determines a set of recognition options for the input grapheme image, wherein the set of recognition options comprises a set of target characters that are similar to the input grapheme image, selects a first neural network from the plurality of neural networks, wherein the first neural network is trained to recognize a first set of confused graphemes, and wherein the first set of graphemes comprises at least a portion of the set of recognition options for the input grapheme image, and determines a grapheme class for the input grapheme image using the selected first neural network.

Abstract: Disclosed are an apparatus and method for detecting a debatable document. According to an embodiment of the present disclosure, the method for detecting a debatable document includes the steps of receiving a document including one or more sentences; generating an embedding vector for each of words included in the document; and extracting features of the document from an embedding vector matrix including the embedding vectors for the words, and detecting debatability of the document from the extracted features through a detection model including a two-step convolutional neural network.

Abstract: Methods and systems are provided for deblurring images. A neural network is trained where the training includes selecting a central training image from a sequence of blurred images. An earlier training image and a later training image are selected based on the earlier training image preceding the central training image in the sequence and the later training image following the central training image in the sequence and based on proximity of the images to the central training image in the sequence. A training output image is generated by the neural network from the central training image, the earlier training image, and the later training image. Similarity is evaluated between the training output image and a reference image. The neural network is modified based on the evaluated similarity. The trained neural network is used to generate a deblurred output image from a blurry input image.

Abstract: Disclosed is a process of providing social networking entity purchasing processes. A method includes receiving, from a posting entity and at the social networking entity, a posting. When the posting is associated with a product within a product catalog of the posting entity, the social networking entity transmits the posting through the social networking entity with an option to buy. When there is a correlation between the posting and the product catalog, and when the user makes a purchase of the product, the user is not transitioned away from the social networking entity. Initiating a process associated with the purchase of the product occurs within the social networking entity.

Abstract: Disclosed is an approach for providing a multi-site shopping cart. A method includes receiving, via a browser shopping cart application programming interface associated with a browser, first data associated with a first product viewed by a user navigating on a first site using the browser. The first data is stored for later retrieval. The method includes presenting a browser payment interface to the user for managing the purchase of the second product, the browser payment interface being associated with a browser payment request API in which payment data for the user is passed from the browser to the second site through the browser payment request application programming interface. The method includes presenting on the browser payment interface information about the first product (based on the stored data) and processing a payment of both the first product and the second product based on user interaction with the browser payment interface.

Abstract: A CNN-based method for meta learning, i.e., learning to learning, by using a learning device including convolutional layers capable of applying convolution operations to an image or its corresponding input feature maps to generate output feature maps, and residual networks capable of feed-forwarding the image or its corresponding input feature maps to next convolutional layer through bypassing the convolutional layers or its sub-convolutional layers is provided. The CNN-based method includes steps of: the learning device (a) selecting a specific residual network to be dropped out among the residual networks; (b) feeding the image into a transformed CNN where the specific residual network is dropped out, and outputting a CNN output; and (c) calculating losses by using the CNN output and its corresponding GT, and adjusting parameters of the transformed CNN. Further, the CNN-based method can be also applied to layer-wise dropout, stochastic ensemble, virtual driving, and the like.

Abstract: Disclosed is an approach for processing cryptocurrency payments via a payment request application programming interface. A method includes receiving, from a site, at a browser and via the payment request application programming interface, a request associated with a potential purchase, wherein the request includes an identification of a cryptocurrency payment method accepted by the site and transmitting, to the site, from the browser and via the API, data indicating that a user of the browser can pay for the potential purchase via the cryptocurrency payment method accepted by the site. The method includes retrieving via the API cryptocurrency payment information for the potential purchase and populating a cryptocurrency wallet with the cryptocurrency payment information for automatic payment.

Abstract: A neural network including an embedding layer to receive a gate function vector and an embedding width and alter a shape of the gate function vector by the embedding width, a concatenator to receive a gate feature input vector and concatenate the gate feature input vector with the gate function vector altered by the embedding width, a convolution layer to receive a window size, stride, and output feature size and generate an output convolution vector with a shape based on a length of the gate function vector, the window size of the convolution layer, and the output feature size of the convolution layer, and a fully connected layer to reduce the gate output convolution vector to a final path delay output.

Abstract: Systems and methods for test object classification are provided in which the test object is docked with a target object in a plurality of different poses to form voxel maps. The maps are vectorized and fed into a convolutional neural network comprising an input layer, a plurality of individually weighted convolutional layers, and an output scorer. The convolutional layers include initial and final layers. Responsive to vectorized input, the input layer feeds values into the initial convolutional layer. Each respective convolutional layer, other than the final convolutional layer, feeds intermediate values as a function of the weights and input values of the respective layer into another of the convolutional layers. The final convolutional layer feeds values into one or more fully connected layers as a function of the final layer weights and input values. The one or more full connected layers feed values into the scorer which scores each input vector to thereby classify the test object.

Abstract: Supervised machine learning using convolutional neural network (CNN) is applied to denoising images rendered by MC path tracing. The input image data may include pixel color and its variance, as well as a set of auxiliary buffers that encode scene information (e.g., surface normal, albedo, depth, and their corresponding variances). In some embodiments, a CNN directly predicts the final denoised pixel value as a highly non-linear combination of the input features. In some other embodiments, a kernel-prediction neural network uses a CNN to estimate the local weighting kernels, which are used to compute each denoised pixel from its neighbors. In some embodiments, the input image can be decomposed into diffuse and specular components. The diffuse and specular components are then independently preprocessed, filtered, and postprocessed, before recombining them to obtain a final denoised image.

Abstract: A method for cropping photos images captured by a user from an image of a page of a photo album is described. Corners in the page image are detected using corner detection algorithm or by detecting intersections of line-segments (and their extensions) in the image using edge, corner, or line detection techniques. Pairs of the detected corners are used to define all potential quads, which are then are qualified according to various criteria. A correlation matrix is generated for each potential pair of the qualified quads, and candidate quads are selected based on the Eigenvector of the correlation matrix. The content of the selected quads is checked using a salience map that may be based on a trained neuron network, and the resulting photos images are extracted as individual files for further handling or manipulation by the user.

Abstract: A method for learning parameters of an object detector based on a CNN adaptable to customers' requirements such as KPI by using a target object merging network and a target region estimating network is provided. The CNN can be redesigned when scales of objects change as a focal length or a resolution changes depending on the KPI. The method includes steps of: a learning device (i) instructing the target region estimating network to search for k-th estimated target regions, (ii) instructing an RPN to generate (k_1)-st to (k_n)-th object proposals, corresponding to an object on a (k_1)-st to a (k_n)-th manipulated images, and (iii) instructing the target object merging network to merge the object proposals and merge (k_1)-st to (k_n)-th object detection information, outputted from an FC layer. The method can be useful for multi-camera, SVM (surround view monitor), and the like, as accuracy of 2D bounding boxes improves.

Abstract: A method for learning parameters of an object detector based on a CNN adaptable to customers' requirements such as KPI by using an image concatenation and a target object merging network is provided. The CNN can be redesigned when scales of objects change as a focal length or a resolution changes depending on the KPI. The method includes steps of: a learning device instructing an image-manipulating network to generate n manipulated images; instructing an RPN to generate first to n-th object proposals respectively in the manipulated images, and instructing an FC layer to generate first to n-th object detection information; and instructing the target object merging network to merge the object proposals and merge the object detection information. In this method, the object proposals can be generated by using lidar. The method can be useful for multi-camera, SVM(surround view monitor), and the like, as accuracy of 2D bounding boxes improves.

Abstract: A method and system of neural network training for mobile device indoor navigation and positioning.

Abstract: A method for learning parameters of a CNN for image recognition is provided to be used for hardware optimization which satisfies KPI. The method includes steps of: a learning device (a) instructing a first transposing layer or a pooling layer to generate an integrated feature map by concatenating pixels, per each ROI, on pooled ROI feature maps; (b) instructing a 1×H1 convolutional layer to generate a first adjusted feature map using a first reshaped feature map, generated by concatenating features in H1 channels of the integrated feature map, and instructing a 1×H2 convolutional layer to generate a second adjusted feature map using a second reshaped feature map, generated by concatenating features in H2 channels of the first adjusted feature map; and (c) instructing a second transposing layer or a classifying layer to divide the second adjusted feature map by each pixel, to thereby generate pixel-wise feature maps.

Abstract: A method for learning parameters of an object detector by using a target object estimating network adaptable to customers' requirements such as KPI is provided. When a focal length or a resolution changes depending on the KPI, scales of objects also change. In this method for customer optimizable design, unsecure objects such as falling or fallen objects may be detected more accurately, and also fluctuations of the objects may be detected. Therefore, the method can be usefully performed for military purpose or for detection of the objects at distance. The method includes steps of: a learning device instructing an RPN to generate k-th object proposals on k-th manipulated images which correspond to (k?1)-th target region on an image; instructing an FC layer to generate object detection information corresponding to k-th objects; and instructing an FC loss layer to generate FC losses, by increasing k from 1 to n.

Abstract: The methods and apparatuses described herein generally relate to providing a platform for allowing a website user to select a product for purchase at a non-merchant website. For example, a commerce engine can receive a request to review product information from a non-merchant website, and can translate the request into a format that can be understood by at least one merchant server using at least one type of commerce platform. The commerce engine can send the translated requests to at least one merchant server, and the merchant servers that receive the requests can determine information about the product (e.g., remaining inventory at particular merchants, product price, and/or other product details). The merchant servers can provide this information to the commerce engine, which can send the product information to the non-merchant website. The commerce engine can also facilitate a transaction with the merchant server, based on the product information returned by the non-merchant website.

Abstract: A method for learning parameters of an object detector based on a CNN is provided to be used for hardware optimization which satisfies KPI. The method includes steps of: a learning device instructing a first transposing layer or a pooling layer to generate an integrated feature map by concatenating pixels per each proposal; and instructing a second transposing layer or a classifying layer to divide volume-adjusted feature map, generated by using the integrated feature map, by pixel, and instructing the classifying layer to generate object class information. By this method, size of a chip can be decreased as convolution operations and fully connected layer operations can be performed by a same processor. Accordingly, there are advantages such as no need to build additional lines in a semiconductor manufacturing process, power saving, more space to place other modules instead of an FC module in a die, and the like.

Abstract: A method for learning parameters of an object detector based on a CNN is provided to be used for hardware optimization which satisfies KPI. The method includes steps of: a learning device (a) instructing a first transposing layer or a pooling layer to concatenate pixels, per each proposal, on pooled feature maps per each proposal; (b) instructing a 1×H1 and a 1×H2 convolutional layers to apply a 1×H1 and a 1×H2 convolution operations to reshaped feature maps generated by concatenating each feature in each of corresponding channels among all channels of the concatenated pooled feature map, to thereby generate an adjusted feature map; and (c) instructing a second transposing layer or a classifying layer to generate pixel-wise feature maps per each proposal by dividing the adjusted feature map by each pixel, and backpropagating object detection losses calculated by referring to object detection information and its corresponding GT.