Techniques are described related to output and removal of decoded pictures from a decoded picture buffer (DPB). The example techniques may remove a decoded picture from the DPB prior to coding a current picture. For instance, the example techniques may remove the decoded picture if that decoded picture is not identified in the reference picture set of the current picture.

A method of determining a reference picture set (RPS), which is a set of reference pictures used in predictive decoding of a current picture that is to be decoded includes: obtaining a flag indicating whether the RPS is determined based on picture order count (POC) values of the current picture and a previous picture or whether the RPS is determined based on an index of a reference RPS, which is an identification value of the reference RPS that is one of pre-defined RPSs and is referred to in determining the RPS, and a delta RPS that is a difference value between a POC vale of a reference picture included in the reference RPS and a POC value of a reference picture included in the RPS; and determining the RPS according to a value of the flag.

A method of determining a reference picture set (RPS), which is a set of reference pictures used in predictive decoding of a current picture that is to be decoded includes: obtaining a flag indicating whether the RPS is determined based on picture order count (POC) values of the current picture and a previous picture or whether the RPS is determined based on an index of a reference RPS, which is an identification value of the reference RPS that is one of pre-defined RPSs and is referred to in determining the RPS, and a delta RPS that is a difference value between a POC vale of a reference picture included in the reference RPS and a POC value of a reference picture included in the RPS; and determining the RPS according to a value of the flag.

A method of coding video data, including constructing a history-based motion vector prediction (HMVP) candidate history table that includes motion vector information of previously coded blocks that extend beyond adjacent neighboring blocks of a current block, constructing a motion vector predictor list, and adding one or more HMVP candidates from the HMVP candidate history table to the motion vector predictor list. Adding the one or more HMVP candidates from the HMVP candidate history table comprises comparing a first HMVP candidate in the HMVP candidate history table to two entries in the motion vector predictor list and no other entries, and adding the first HMVP candidate to the motion vector predictor list when the first HMVP candidate is different than both of the two entries in the motion vector predictor list. The method also includes coding the current block of video data using the motion vector predictor list.

A method of coding video data, including constructing a history-based motion vector prediction (HMVP) candidate history table that includes motion vector information of previously coded blocks that extend beyond adjacent neighboring blocks of a current block, constructing a motion vector predictor list, and adding one or more HMVP candidates from the HMVP candidate history table to the motion vector predictor list. Adding the one or more HMVP candidates from the HMVP candidate history table comprises comparing a first HMVP candidate in the HMVP candidate history table to two entries in the motion vector predictor list and no other entries, and adding the first HMVP candidate to the motion vector predictor list when the first HMVP candidate is different than both of the two entries in the motion vector predictor list. The method also includes coding the current block of video data using the motion vector predictor list.

A target identification method includes receiving target detection information for a current target from a device, and acquiring target state information of multiple reference targets. A degree of matching between the target detection information and the target state information of each of the multiple reference targets is determined; and the current target is associated with a corresponding reference target having a best degree of matching amongst the multiple reference targets.

A target identification method includes receiving target detection information for a current target from a device, and acquiring target state information of multiple reference targets. A degree of matching between the target detection information and the target state information of each of the multiple reference targets is determined; and the current target is associated with a corresponding reference target having a best degree of matching amongst the multiple reference targets.

A decoder includes circuitry configured to receive a bitstream; store a plurality of long-term reference frames in a reference list; retain a long-term reference frame in the reference list for a length of time based on a retention time; and decode at least a portion of video using the long-term reference frame retained in the reference list. Related apparatus, systems, techniques and articles are also described.

This disclosure relates encoding and decoding of motion vector difference in for inter-predicting a video block. An example is disclosed for decoding an inter-predicted video block of a video stream. The method includes receiving a coded video bitstream; extracting, from the coded video bitstream, an inter-predication mode and a joint delta motion vector (MV) for a current block in a current frame; extracting, from the coded video bitstream, a flag indicating whether a first delta MV for a first reference frame and a second delta MV for a second reference frame are jointly signaled; in response to the flag indicating that the first delta MV and the second delta MV are jointly signaled, deriving the first delta MV and the second delta MV based on the joint delta MV; and decoding the current block in the current frame based on the first delta MV and the second delta MV.

A visual media processing method includes determining that a first video block of a visual media data uses a geometric partitioning mode (GPM) and a second video block of the visual media data uses a non-GPM mode; constructing, based on a unified pruning process, a first merge list for the first video block and a second merge list for the second video block, wherein the first merge list and the second merge list comprise merge candidates, and the pruning process includes adding a new merge candidate to the merge list based on comparing motion information of the new merge candidate with motion information of at least one merge candidate in the merge list, wherein the GPM comprises splitting the first video block into multiple prediction partitions to apply motion prediction separately, and at least one partition having a non-rectangular shape.