Abstract: Embodiments are disclosed for generating temporally consistent manipulated videos. A method of generating temporally consistent manipulated videos comprises receiving a target appearance and an input digital video including a plurality of frames, generating a plurality of target appearance frames from the plurality of frames, training a video prediction network to generate a digital video wherein a subject of the digital video has its appearance modified to match the target appearance, providing the input digital video to the video prediction network, and generating, by the video prediction network, an output digital video wherein the subject of the output digital video has its appearance modified to match the target appearance.

Abstract: Embodiments of the present invention are directed towards reframing videos from one aspect ratio to another aspect ratio while maintaining visibility of regions of interest. A set of regions of interest are determined in frames in a video with a first aspect ratio. The set of regions of interest can be used to estimate an initial camera path. An optimal camera path is determined by leveraging the identified regions of interest using the initial camera path. Sub crops with a second aspect ratio different from the first aspect ratio of the video are identified. The sub crops are placed as designated using the optimal camera path to generate a cropped video with the second aspect ratio.

Abstract: This disclosure describes one or more implementations of a video inference system that utilizes machine-learning models to efficiently and flexibly process digital videos utilizing various improved video inference architectures. For example, the video inference system provides a framework for improving digital video processing by increasing the efficiency of both central processing units (CPUs) and graphics processing units (GPUs). In one example, the video inference system utilizes a first video inference architecture to reduce the number of computing resources needed to inference digital videos by analyzing multiple digital videos utilizing sets of CPU/GPU containers along with parallel pipeline processing. In a further example, the video inference system utilizes a second video inference architecture that facilitates multiple CPUs to preprocess multiple digital videos in parallel as well as a GPU to continuously, sequentially, and efficiently inference each of the digital videos.

Abstract: Embodiments of the present invention are directed towards reframing videos from one aspect ratio to another aspect ratio while maintaining visibility of regions of interest. A set of regions of interest are determined in frames in a video with a first aspect ratio. The set of regions of interest can be used to estimate an initial camera path. An optimal camera path is determined by leveraging the identified regions of interest using the initial camera path. Sub crops with a second aspect ratio different from the first aspect ratio of the video are identified. The sub crops are placed as designated using the optimal camera path to generate a cropped video with the second aspect ratio.

Abstract: Embodiments of the present invention are directed towards reframing videos from one aspect ratio to another aspect ratio while maintaining visibility of regions of interest. A set of regions of interest are determined in frames in a video with a first aspect ratio. The set of regions of interest can be used to estimate an initial camera path. An optimal camera path is determined by leveraging the identified regions of interest using the initial camera path. Sub crops with a second aspect ratio different from the first aspect ratio of the video are identified. The sub crops are placed as designated using the optimal camera path to generate a cropped video with the second aspect ratio.

Abstract: Disclosed herein are techniques for detecting logos in images or video. In one embodiment, one or more candidate regions are detected for determining logos in an image. A logo is determined to be the logo in the candidate region based on matching a feature vector of a candidate region to a feature vector of the logo.

Abstract: Disclosed herein are techniques for detecting logos in images or video. In one embodiment, one or more candidate regions are detected for determining logos in an image. A logo is determined to be the logo in the candidate region based on matching a feature vector of a candidate region to a feature vector of the logo.

Abstract: Embodiments of the present invention are directed towards reframing videos from one aspect ratio to another aspect ratio while maintaining visibility of regions of interest. A set of regions of interest are determined in frames in a video with a first aspect ratio. The set of regions of interest can be used to estimate an initial camera path. An optimal camera path is determined by leveraging the identified regions of interest using the initial camera path. Sub crops with a second aspect ratio different from the first aspect ratio of the video are identified. The sub crops are placed as designated using the optimal camera path to generate a cropped video with the second aspect ratio.

Abstract: Disclosed herein are techniques for detecting logos in images or video. In one embodiment, a first logo detection model detects, from an image, candidate regions for determining logos in the image. A feature vector is then extracted from each candidate region and is compared with reference feature vectors stored in a database. The logo corresponding to the best matching reference feature vector is determined to be the logo in the candidate region if the best matching meets a certain criterion. In some embodiments, a second logo detection model trained using synthetic training images is used in combination with the first logo detection model to detect logos in a same image.

Abstract: Disclosed herein are techniques for determining brand safety of a video including image frames and audio content. In some embodiments, frame-level features, scene-level features, and video-level features are extracted by a set of frame-level models, a set of scene-level models, and a set of video-level models, respectively. Outputs from lower level models are used as inputs for higher level models. A brand safety score indicating whether it is safe to associate a brand with the video is determined based on the outputs from the set of video-level models. In some embodiments, commercial content associated with the brand is insert into the video that is determined to be safe for the brand.

Abstract: Techniques disclosed herein relate generally to augmenting images or videos with graphics. More specifically, some embodiments relate to realistically or photorealistically augmenting a target image or video frame with a source graph, such as a computer-generated graph or a real world image. In one embodiment, a planar segment of the target image is identified based on a surface normal map of the target image. The planar segment is then used to determine a focal length and a homography function for transforming the source graph.

Abstract: Disclosed herein are techniques for detecting logos in images or video. In one embodiment, a first logo detection model detects, from an image, candidate regions for determining logos in the image. A feature vector is then extracted from each candidate region and is compared with reference feature vectors stored in a database. The logo corresponding to the best matching reference feature vector is determined to be the logo in the candidate region if the best matching meets a certain criterion. In some embodiments, a second logo detection model trained using synthetic training images is used in combination with the first logo detection model to detect logos in a same image.

Abstract: Disclosed herein are techniques for determining brand safety of a video including image frames and audio content. In some embodiments, frame-level features, scene-level features, and video-level features are extracted by a set of frame-level models, a set of scene-level models, and a set of video-level models, respectively. Outputs from lower level models are used as inputs for higher level models. A brand safety score indicating whether it is safe to associate a brand with the video is determined based on the outputs from the set of video-level models. In some embodiments, commercial content associated with the brand is insert into the video that is determined to be safe for the brand.

Abstract: Disclosed herein are techniques for digital content integration. A computer-implemented method includes receiving a target digital content item that includes a plurality of frames, identifying a set of candidate host frames for inserting source digital content items from the plurality of frames based on one or more attributes of the target digital content item, determining a candidate score for each respective candidate host frame of the candidate host frames, and generating host time defining data including identifications and the candidate scores of the candidate host frames, where the candidate score indicates a degree of transition of the target digital content item at the candidate host frame. One or more candidate host frames are then selected based on the candidate scores for inserting one or more source digital content items into the target digital content item.

Abstract: Techniques disclosed herein relate generally to augmenting images or videos with graphics. More specifically, some embodiments relate to realistically or photorealistically augmenting a target image or video frame with a source graph, such as a computer-generated graph or a real world image. In one embodiment, a planar segment of the target image is identified based on a surface normal map of the target image. The planar segment is then used to determine a focal length and a homography function for transforming the source graph.

Abstract: Disclosed herein are techniques for digital content integration. A computer-implemented method includes receiving a target digital content item that includes a plurality of frames, identifying a set of candidate host frames for inserting source digital content items from the plurality of frames based on one or more attributes of the target digital content item, determining a candidate score for each respective candidate host frame of the candidate host frames, and generating host time defining data including identifications and the candidate scores of the candidate host frames, where the candidate score indicates a degree of transition of the target digital content item at the candidate host frame. One or more candidate host frames are then selected based on the candidate scores for inserting one or more source digital content items into the target digital content item.