JVET-AD0129 AHG10: Improvement of Input Video Padding in VTM [J. Liao, L. Li, D. Liu, H. Li, F. Wu (USTC)]

This contribution proposes to fix the input video padding implementation in VTM software. In order to handle inputs of various sizes, VTM pads the input video. In the original padding implementation, compared with the normal replicate padding in the bottom side, the pixels in the right side are padded with the upper right pixel without any sense. This contribution proposes to fix the right-side padding as a replicate padding.

Since the normal size of CTC sequences does not trigger the implementation of padding, this contribution tests the performance on CTC sequences cropped with a reduction of 4 in both original width and height. Moreover, this contribution demonstrates the test results on CTC sequences cropped respectively with the center point (CT) and upper left corner (UL) as the center.

The test results show that the fix brings significant performance improvement compared with the original implementation as follows:

CT:

AI: {-0.17%, -0.41%, -0.36%}; RA: {-0.33%, -0.72%, -0.64%};

LDB: {-0.31%, -0.62%, -0.63%}

UL:

AI: {-0.16%, -0.41%, -0.45%}; RA: {-0.36%, -0.65%, -0.31%};

LDB: {-0.28%, -0.39%, -0.74%}

This is related to padding the input pictures to a size that can be handled by VVC. This condition is not triggered by our CTC sequences.

It was commented that class C shows some consistent luma loss for LD and RA. The proponent suggests that this is mainly due to the RaceHorse sequence which has moving objects along the rightside border of the picture.

A cross checker confirmed the results and the implementation.

It was commented that this looks like an obvious bug, and could be taken into the VTM after the VTM software coordinators perform a code review and confirm the implementation. This was later confirmed by X. Li.

Decision (SW): Adopt JVET-AD0129