Smoothing overlapped regions resulting from geometric motion partitioning

摘要

In one example, an apparatus includes a video encoder configured to partition a block of video data into a first partition and a second partition using a geometric motion partition line, calculate a prediction value of a pixel in a transition region of the block using a filter that applies a value for at least one neighboring pixel from the first partition and a value for at least one neighboring pixel from the second partition, calculate a residual value of the pixel in the transition region of the block based on the prediction value of the pixel in the transition region, and output the residual value of the pixel. In one example, a video decoder may use a similar filter to decode an the encoded block after receiving the residual value for the encoded block, and using a definition of the geometric motion partition line.

主权项

1. A method for encoding video data, the method comprising:
partitioning a block of video data into a first partition and a second partition using a geometric motion partition line; determining whether to use a filter that applies a value for at least one neighboring pixel from the first partition and a value for at least one neighboring pixel from the second partition during calculation of a prediction value of a pixel in a transition region of the block based on one or more conditions related to the pixel in the transition region, wherein the one or more conditions include at least one of a difference of pixel values along a line that is perpendicular to the geometric motion partition line, a difference between a motion vector for the first partition and a motion vector for the second partition, a direction of the geometric motion partition line, reference frame indexes for the first partition and the second partition, quantization parameters for the first partition and the second partition, and numbers of pixels in the first partition and the second partition; in response to determining to use the filter, calculating the prediction value of the pixel using the filter, wherein calculating the prediction value comprises determining weights to apply to the value for the at least one neighboring pixel from the first partition and to the value for the at least one neighboring pixel from the second partition based on one or more conditions related to the pixel in the transition region, wherein the one or more conditions include at least one of the difference of pixel values along the line that is perpendicular to the geometric motion partition line, the difference between the motion vector for the first partition and the motion vector for the second partition, the direction of the geometric motion partition line, the reference frame indexes for the first partition and the second partition, the quantization parameters for the first partition and the second partition, and the numbers of pixels in the first partition and the second partition; calculating a residual value of the pixel in the transition region of the block based on the prediction value of the pixel in the transition region; and outputting the residual value of the pixel.