JVET-AD0138 [AHG8] QPA with low activity threshold for machine task [C. Kim, D. Gwak, J. Lim (LGE)]

This contribution proposes a low activity threshold for QPA to reduce bit rate increases in less important areas for machine task. From experimental results, 1.19% performance improvement is observed on average.

At the previous JVET meeting, the impact of the perceptual QP adaptation (QPA) on object detection task performance for Open Images dataset was reported JVET-AC0079. This contribution reports on the object detection performance of the perceptual QP adaptation (QPA) on the SFU-HW video dataset, as well as the object detection performance using QPA with a low activity threshold.

Average performance improvement of 1.19% is reported compared to QPA which is in the software repository.

It is however commented that the results are inhomogeneous: In some cases, the results are better compared to current QPA, in other cases worse. Fluctuations are significant.

Further study was recommended to make the improvement more homogeneous.

It was also commented that BD rate changes in the range of 1% are not indicating much in this activity.