Abstract: Methods, system, and computer storage media are provided for novel view synthesis. An input image depicting an object is received and utilized to generate, via a neural network, a target view image. In exemplary aspects, additional view images are also generated within the same pass of the neural network. A loss is determined based on the target view image and additional view images and is used to modify the neural network to reduce errors. In some aspects, a rotated view image is generated by warping a ground truth image from an initial angle to a rotated view angle that matches a view angle of an image synthesized via the neural network, such as a target view image. The rotated view image and the synthesized image matching the rotated view angle (e.g., a target view image) are utilized to compute a rotational loss.

Abstract: Systems and methods predict a performance metric for a video and identify key portions of the video that contribute to the performance metric, which can be used to edit the video to improve the ultimate viewer response to the video. An initial performance metric is computed for an initial video (e.g., using a neural network). A perturbed video is generated by perturbing a video portion of the initial video. A modified performance metric is computed for the perturbed video. Based on a difference between the initial and modified performance metrics, the system determines that the video portion contributed to a predicted user viewer response to the initial video. An indication of the video portion that contributed to the predicted user viewer response is provided as output, which can be used to edit the video to improve the predicted viewer response.

Abstract: A video summarization system generates a concatenated feature set by combining a feature set of a candidate video shot and a summarization feature set. Based on the concatenated feature set, the video summarization system calculates multiple action options of a reward function included in a trained reinforcement learning module. The video summarization system determines a reward outcome included in the multiple action options. The video summarization system modifies the summarization feature set to include the feature set of the candidate video shot by applying a particular modification indicated by the reward outcome. The video summarization system identifies video frames associated with the modified summarization feature set, and generates a summary video based on the identified video frames.

Abstract: A video summarization system generates a concatenated feature set by combining a feature set of a candidate video shot and a summarization feature set. Based on the concatenated feature set, the video summarization system calculates multiple action options of a reward function included in a trained reinforcement learning module. The video summarization system determines a reward outcome included in the multiple action options. The video summarization system modifies the summarization feature set to include the feature set of the candidate video shot by applying a particular modification indicated by the reward outcome. The video summarization system identifies video frames associated with the modified summarization feature set, and generates a summary video based on the identified video frames.

Abstract: This disclosure generally covers systems and methods that determine demographic labels for a user or a group of users by using digital inputs within a predictive model for demographic classification. In particular, the disclosed systems and methods use a unique combination of classification algorithms to determine demographic labels for users as a potential audience of digital content items. When applying the combination of classification algorithms, the disclosed systems and methods use a first classification algorithm to determine user-level-latent features for each user within a group of users based on demographic-label statistics associated with particular digital content items. The disclosed systems and methods then use the user-level-latent features and session-level features (from sessions of each user consuming the digital content items) as inputs in a second classification algorithm to determine a demographic label for each user within the group of users.

Abstract: Generating images and videos depicting a human subject wearing textually defined attire is described. An image generation system receives a two-dimensional reference image depicting a person and a textual description describing target clothing in which the person is to be depicted as wearing. To maintain a personal identity of the person, the image generation system implements a generative model, trained using both discriminator loss and perceptual quality loss, which is configured to generate images from text. In some implementations, the image generation system is configured to train the generative model to output visually realistic images depicting the human subject in the target clothing. The image generation system is further configured to apply the trained generative model to process individual frames of a reference video depicting a person and output frames depicting the person wearing textually described target clothing.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods to generate sketches for clearing-bid values and bid-success rates based on multi-dimensional targeting criteria for a digital-content campaign and dynamically determine predicted values for the digital-content campaign based on the sketches. To illustrate, the disclosed systems can use a running-average-tuple-sketch to generate tuple sketches of historical clearing-bid values and tuple sketches of historical bid-success-rates from historical auction data. Based on the tuple sketches, the disclosed systems can determine one or more of a predicted cost per quantity of impressions, a predicted number of impressions, or a predicted expenditure for the digital-content campaign—according to user-input targeting criteria and expenditure constraints.

Abstract: In implementations of trajectory-based viewport prediction for 360-degree videos, a video system obtains trajectories of angles of users who have previously viewed a 360-degree video. The angles are used to determine viewports of the 360-degree video, and may include trajectories for a yaw angle, a pitch angle, and a roll angle of a user recorded as the user views the 360-degree video. The video system clusters the trajectories of angles into trajectory clusters, and for each trajectory cluster determines a trend trajectory. When a new user views the 360-degree video, the video system compares trajectories of angles of the new user to the trend trajectories, and selects trend trajectories for a yaw angle, a pitch angle, and a roll angle for the user. Using the selected trend trajectories, the video system predicts viewports of the 360-degree video for the user for future times.

Abstract: A baseboard management controller (BMC) may be configured to enable a communication interface from the BMC to a host processor on a host computing device and provide input to the host processor via the communication interface. The input causes at least one diagnostic operation to be performed on the host computing device. The BMC may collect host information in response to the diagnostic operation(s) being performed. The BMC may report the host information to another entity and/or store the host information in persistent memory within the BMC. In some embodiments, the input may be provided to the host processor in response to receiving a signal from a fabric controller. In some embodiments, the input may be provided to the host processor in response to detecting an anomaly associated with the host computing device. The BMC may take at least one action to mitigate the anomaly.

Abstract: A deep relational factorization machine (“DRFM”) system is configured to provide a high-order prediction based on high-order feature interaction data for a dataset of sample nodes. The DRFM system can be configured with improved factorization machine (“FM”) techniques for determining high-order feature interaction data describing interactions among three or more features. The DRFM system can be configured with improved graph convolutional neural network (“GCN”) techniques for determining sample interaction data describing sample interactions among sample nodes, including sample interaction data that is based on the high-order feature interaction data. The DRFM system generates a high-order prediction based on the high-order feature interaction embedding vector and the sample interaction embedding vector. The high-order prediction can be provided to a prediction computing system configured to perform operations based on the high-order prediction.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods that iteratively select versions of augmented reality objects at augmented reality levels of detail to provide for download to a client device to reduce start-up latency associated with providing a requested augmented reality scene. In particular, in one or more embodiments, the disclosed systems determine utility and priority metrics associated with versions of augmented reality objects associated with a requested augmented reality scene. The disclosed systems utilize the determined metrics to select versions of augmented reality objects that are likely to be viewed by the client device and improve the quality of the augmented reality scene as the client device moves through the augmented reality scene. In at least one embodiment, the disclosed systems iteratively select versions of augmented reality objects at various levels of detail until the augmented reality scene is fully downloaded.

Abstract: Generating images and videos depicting a human subject wearing textually defined attire is described. An image generation system receives a two-dimensional reference image depicting a person and a textual description describing target clothing in which the person is to be depicted as wearing. To maintain a personal identity of the person, the image generation system implements a generative model, trained using both discriminator loss and perceptual quality loss, which is configured to generate images from text. In some implementations, the image generation system is configured to train the generative model to output visually realistic images depicting the human subject in the target clothing. The image generation system is further configured to apply the trained generative model to process individual frames of a reference video depicting a person and output frames depicting the person wearing textually described target clothing.

Abstract: Techniques are disclosed for the improvement of vector quantization (VQ) codebook generation. The improved codebooks may be used for compression in cloud-based video applications. VQ achieves compression by vectorizing input video streams, matching those vectors to codebook vector entries, and replacing them with indexes of the matched codebook vectors along with residual vectors to represent the difference between the input stream vector and the codebook vector. The combination of index and residual is generally smaller than the input stream vector which they collectively encode, thus providing compression. The improved codebook may be generated from training video streams by grouping together similar types of data (e.g., image data, motion data, control data) from the video stream to generate longer vectors having higher dimensions and greater structure. This improves the ability of VQ to remove redundancy and thus increase compression efficiency.

Abstract: Techniques are disclosed for improving media content effectiveness. A methodology implementing the techniques according to an embodiment includes generating an intermediate representation (IR) of provided media content, the IR specifying editable elements of the content and maintaining a result of cumulative edits to those elements. The method also includes editing the elements of the IR to generate a set of candidate IR variations. The method further includes creating a set of candidate media contents based on the candidate IR variations, evaluating the candidate media contents to generate effectiveness scores, and pruning the set of candidate IR variations to retain a threshold number of the candidate IR variations as surviving IR variations associated with the highest effectiveness scores. The process iterates until either an effectiveness score exceeds a threshold value, the incremental improvement at each iteration falls below a desired value, or a maximum number of iterations have been performed.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer-readable media to enhance texture image delivery and processing at a client device. For example, the disclosed systems can utilize a server-side compression combination that includes, in sequential order, a first compression pass, a decompression pass, and a second compression pass. By applying this compression combination to a texture image at the server-side, the disclosed systems can leverage both GPU-friendly and network-friendly image formats. For example, at a client device, the disclosed system can instruct the client device to execute a combination of decompression-compression passes on a GPU-network-friendly image delivered over a network connection to the client device.

Abstract: Embodiments of the present disclosure provide systems, methods, and computer storage media for optimizing computing resources generally associated with cloud-based media services. Instead of decoding digital assets on-premises to stream to a remote client device, an encoded asset can be streamed to the remote client device. A codebook employable for decoding the encoded asset can be embedded into the stream transmitted to the remote client device, so that the remote client device can extract the embedded codebook, and employ the extracted codebook to decode the encoded asset locally. In this way, not only are processing resources associated with on-premises decoding eliminated, but on-premises storage of codebooks can be significantly reduced, while expensive bandwidth is freed up by virtue of transmitting a smaller quantity of data from the cloud to the remote client device.

Abstract: A baseboard management controller (BMC) may comprise a processor, a non-volatile memory and a volatile memory. The non-volatile memory comprises firmware categorized into a plurality of independently updatable service modules. Each of the independently updatable service modules is stored on a read-write partition of the non-volatile memory and comprises at least one of an application, a library and a driver. The BMC comprises an update agent that performs an update process. In the update process, a BMC update package, which comprises an update service module for updating an existing service module stored in one of the plurality of RW partitions, is stored in the volatile memory. The existing service module stored in the RW partition is replaced with the update service module.

Abstract: The present disclosure relates to systems, non-transitory computer-readable media, and methods that iteratively select versions of augmented reality objects at augmented reality levels of detail to provide for download to a client device to reduce start-up latency associated with providing a requested augmented reality scene. In particular, in one or more embodiments, the disclosed systems determine utility and priority metrics associated with versions of augmented reality objects associated with a requested augmented reality scene. The disclosed systems utilize the determined metrics to select versions of augmented reality objects that are likely to be viewed by the client device and improve the quality of the augmented reality scene as the client device moves through the augmented reality scene. In at least one embodiment, the disclosed systems iteratively select versions of augmented reality objects at various levels of detail until the augmented reality scene is fully downloaded.

Abstract: Residual vectors are compressed in a lossless compression scheme suitable for cloud DVR video content applications. Thus, a cloud DVR service provider can take many copies of the same file stored in the cloud and save storage space by compressing those copies while still maintaining their status as distinct copies, one per user. Vector quantization is used for compressing already-compressed video streams (e.g., MPEG streams). As vector quantization is a lossy compression scheme, the residual vector has to be stored to regenerate the original video stream at the decoding (playback) node. Entropy coding schemes like Arithmetic or Huffman coding can be used to compress the residual vectors. Additional strategies can be implemented to further optimize this residual compression. In some embodiments, the techniques operate to provide a 25-50% improvement in compression. Storage space is thus more efficiently used and video transmission may be faster in some cases.

Abstract: Embodiments of a system and method for low-latency content streaming are described. In various embodiments, multiple data fragments may be sequentially generated. Each data fragment may represent a distinct portion of media content generated from a live content source. Each data fragment may include multiple sub-portions. Furthermore, for each data fragment, generating that fragment may include sequentially generating each sub-portion of that fragment. Embodiments may include, responsive to receiving a request for a particular data fragment from a client during the generation of a particular sub-portion of that particular data fragment, providing the particular sub-portion to the client subsequent to that particular sub-portion being generated and prior to the generation of that particular data fragment being completed in order to reduce playback latency at the client relative to the live content source.