OP Solutions, LLC

An encoder for encoding video for machine consumption includes a first feature encoding element and a second feature encoding element. The first feature encoding element receives a set of features extracted from a source video by a backbone of a convolutional neural network and outputs features having reduced dimensions compared to the features input into the first feature encoding element. The second feature encoding element receiving the features having reduced dimensions and outputting an encoded bitstream for a machine video application.

A decoder includes circuitry configured to receive a bitstream, identify a first frame, locate a first independent reference region within the first frame, extract the first independent reference region from the first frame, decode a second frame using the first independent reference region as a reference for the second frame.

Systems and methods for encoding and decoding video for machine consumption (video coding for machines) are provided in which an autoencoder is employed. The autoencoder has an encoder portion, a bottleneck portion, and a decoder portion. The autoencoder being distributed between a VCM encoder and VCM decoder such that the VCM encoder includes the encoder portion and bottleneck portion and the VCM decoder includes the bottleneck portion and the decoder portion.

Methods and systems for hybrid feature video bitstream encoding and decoding are provided. Encoding can be performed in one of a plurality of modes. In one mode, the video components of the encoded bitstream include residual data generated encoding the difference between feature data and input video data. A compatible decoder can operate in a plurality of decoding modes. In one decoding mode, the video components are decoded using the encoded video and encoded feature data in the bitstream. The hybrid bitstream between the encoder and decoder is formed with a plurality of hybrid segments, each having a plurality of components specifying characteristics the feature and video content of the segment.

A system and method for video and feature coding of neural-network structures used for multi-task machine learning includes an encoder, decoder and a decoder-compliant bitstream. A task-specific video decoder includes a first decoder receiving a bitstream having at least one feature and a description of a neural network backbone used to generate the bitstream, and a task-specific neural network head. The neural network head recreates the neural network backbone from the description, receives a feature from the bitstream a generates a task-specific output.

A method of memory-efficient prediction mode selection includes receiving, by an encoder, a coded bitstream including a current frame, determining, by the encoder, costs of a first prediction mode and a second prediction mode, wherein determining further comprises determining, for the first prediction mode, a first bit cost and a first memory cost and determining, for the second prediction mode, a second bit cost and a second memory cost, selecting, by the encoder, a current prediction mode of the first prediction mode and the second prediction mode as a function of the first bit cost, first memory cost, second bit cost, and second memory cost, and encoding, by the encoder, the current frame using the current prediction mode. The prediction mode may be informed by at least one parameter received from the decoder.

A decoder includes circuitry configured to receive a bitstream; determine, for a current block and using the bitstream, a global motion vector candidate utilized by an adjacent block; construct a motion vector candidate list including adding the determined global motion vector candidate to the motion vector candidate list; and reconstruct pixel data of the current block and using the motion vector candidate list. Related apparatus, systems, techniques and articles are also described.

A video encoder is configured to generate a bitstream for decoding by a compatible decoder which includes circuitry configured to receive a bitstream, extract a residual of a control point motion vector for a current frame and from the bitstream, and combine the residual of the control point motion vector with a prediction of the control point motion vector for the current frame.

An encoder for generating an encoded bitstream for decoding by a decoder includes a circuitry configured to receive a bitstream including a coded picture having a coding unit with at least one non-straight, non-rectangular boundary therein, in which decoding includes: determining start and end points of the boundary using first and second index information provided in the bitstream, generating predictive pixel values in the regions on each side of the boundary, smoothing the predictive pixel values across the boundary, and adding residual pixel values to the smoothed predictive pixel values.

A video encoder encodes a plurality of pictures with one of the pictures being a long term reference picture. The encoder provides signaling in a bitstream for a decoder to reconstruct pictures by selectively activating the long term reference picture for use as a reference picture, maintaining the long term reference picture in the buffer while being inactive as a reference picture, changing the long term reference picture from inactive to active as a reference picture, and marking the long term reference picture as being unused for reference so that it is no longer eligible to be used as a reference picture for a subsequent picture and is removed as a reference picture.