Amazon Technologies, Inc.

Hai Wei

Yongjun Wu

Lei Li

Charles Benjamin Franklin Waggoner

Abhishek Kumar

Yang Yang

Sriram Sethuraman

Amarsingh Buckthasingh Winston

Charles Benjamin Waggoner

Zongyi Liu

Brian Manuel Alvarez

Srinivas Rajagopalan

Deepthi Nandakumar

Avisar Ten-Ami

Zaixi Shang

Xiaohan Nie

Srikanth Kiran Kotagiri

Bruce Ferry

Dag Henriksson

Ola Hallmarker

Joshua Peter Ebenezer

Simon Paul Lacasse

Yixu Chen

Techniques are described for training and use of machine learning models to determine objective video quality scores. Video quality scores predict the quality of video content perceived by viewers. Quality scores have various uses, including the selection of encoding profiles and determination of encoding ladders. A core model and residual model may be used to determine quality scores.

Objective video quality assessment models based on bitstream, and additional pixel domain features

Techniques are described for facilitating client-side augmented reality overlay of secondary content during live events. Regions for overlaying secondary content are identified along with attributes for each region. A client device may then used the attributes to overlay secondary content in each region prior to playback.

Client side augmented reality overlay

Techniques for emulating frequency coefficients of film grain for video (e.g., frame) encoding are described. According to some embodiments, a computer-implemented method includes receiving a request to encode a video with film grain at a content delivery service, performing a film grain synthesis on a frame of the video with film grain to generate film grain metadata for the frame and generate a degrained frame from the frame, determining film grain emulation frequency coefficients based at least in part on the film grain metadata for the frame, performing a frequency transform on the degrained frame to generate frequency coefficients for the degrained frame, encoding the frame into an encoded frame, by the content delivery service, based at least in part on the frequency coefficients for the degrained frame and the film grain emulation frequency coefficients, and transmitting the encoded frame from the content delivery service to a viewer device.

Efficient coefficient film grain synthesis for codecs without film grain synthesis

Techniques are described for creating and using playback-conditions-adaptive live video encoding ladders.

Content delivery of live streams with playback-conditions-adaptive encoding

Methods and apparatus are described for delivering streams of media content in ways that smooth out the peaks that might otherwise occur due to the bit rate variations that result from encoding of the media content. This is accomplished by controlling the timing of the transmission of packets of the encoded media content.

Smoothing bit rate variations in the distribution of media content

Systems and methods are described herein for detecting image impairments indicating low quality images within video content. An image impairment detection (IMD) model can be generated with a training data set including images that include one or more image impairments. Video content comprising a set of video frames may be obtained and a video frame provided to the image impairment detection model as input. Output may be received from the image impairment detection model indicating that the video frame includes an image impairment artifact indicative of a particular image impairment. One or more operations may be executed based at least in part on the output received.

Techniques for detecting low image quality

Techniques for content-adaptive video sampling for automated video quality monitoring are described. As one example, a computer-implemented method includes receiving a request to train a machine learning model on a training video file comprising at least one labeled defect, performing an encode on the training video file to generate one or more compression features for each compressed frame of the training video file, training the machine learning model to identify a proper subset of candidate defect frames of the training video file based at least in part on the one or more compression features for each compressed frame of the training video file and the at least one labeled defect, receiving an inference request for an input video file, performing an encode on the input video file to generate one or more compression features for each compressed frame of the input video file, generating, by the machine learning model, a proper subset of candidate defect frames of the input video file based at least in part on the one or more compression features for each compressed frame of the input video file, and determining a defect in the input video file based at least in part on the proper subset of candidate defect frames of the input video file.

Content-adaptive video sampling for cost-effective quality monitoring

Techniques are described for detecting desynchronization between an audio stream and a video stream.

Detecting audio-video desyncrhonization

Techniques are described for adaptive encoding of different visual elements in a video frame. Characteristics of visual elements can be determined and used to set encoding parameters for the visual elements. The visual elements can be encoded such that one visual element is encoded differently than another visual element if they have different characteristics.

Visual element encoding parameter tuning

Techniques for a fast approximation of linear light scaling for inputs of non-linear light values are described. As one example, a computer-implemented method includes receiving a request to downscale a plurality of pixels of a single frame of a video file performing, in response to the request to downscale, a lookup in a lookup table for a first input of a first non-linear light value of luminance for a first pixel of the plurality of pixels and a second input of a second non-linear light value of luminance for a second pixel of the plurality of pixels to generate an output of a third single non-linear light value of luminance for a linear light scaling for the first pixel and the second pixel, generating a scaled frame based at least in part on the third single non-linear light value of luminance for the linear light scaling, receiving a request for a manifest for the video file from a client device, generating the manifest for the client device that identifies a scaled video representation that comprises the scaled frame, and sending the manifest to the client device.

Linear light scaling service for non-linear light pixel values

Techniques are described for optimizing event-adaptive live video encoding profiles.

Content delivery of live streams with event-adaptive encoding

An adaptive video compression system may receive video data to be compressed, such as for delivery to a user device by a video streaming service. For example the video data may be an entire video file or a segment of a video file. The adaptive video compression system determines a suitable encoding scheme for compressing the video data. In order to determine the encoding scheme, the video data may be analyzed to extract a plurality of features interests of the video data, which may represent one or more characteristics of the video data. The features may be concatenated and collectively expressed as a feature vector. The feature vector is then used to determine a classification for the video data. Accordingly, an encoding scheme is determined for the video data based on the classification of the video data, and applied to video data to compress the video data.

Growing community of inventors

Seattle, WA, United States of America

Hai Wei