JVET-AD0211 EE1-Related: Combination test of EE1-1.3.5 and multi-scale component of EE1-1.6 [Y. Li, S. Eadie, D. Rusanovskyy, M. Karczewicz (Qualcomm)]

Input document JVET-AD0205, after JVET-AB0164, reports results of the EE1-1.3, with the sub-test 1.3.5 studying convolution decompositions in residual blocks of the architecture and achieves 26% complexity reduction. A multi-scale feature extraction (employing point-wise convolution in parallel to 3x3 convolution) was studied in EE1-1.6, and demonstrated coding gain improvement, comparing to FilterSet1.

In this contribution is presented a combination of EE-1 tests 1.3.5 and 1.6. The multi-scale feature extraction component of test 1.6 was integrated into the backbone residual blocks of EE1-1.3.5.

Comparing to NNVC-4.0 anchor (NN tools OFF), proposed combination provides the following BD-Rate gain for RA and AI, respectively:

Model complexity in kMAC/pixel: 353.9 (frame-wise) of (Luma: 283.21, Chroma: 70.71); 447.9 (block-wise) of (Luma: 358.4, Chroma:89.5)

Float: {-10.33%, -24.26%, -25.31%} and {-8.08%, -21.61%, -22.8%}.
Int16: {-10.34%, -24.07%, -25.24%} and {-8.02%, -22.05%, -23.29%}.

Comparing to NNVVC filter set 1, proposed method has 31% lower complexity and demonstrates the following performance:

Float: {-0.81%, -4.52%, -4.92%} and {-0.75%, -4.21%, -3.6%}.
Int16: {-0.81%, -4.25%, -4.83%} and {-0.67%, -4.71%, -4.15%}.

Proponents proposed to study the proposed method in the next round of EE1.

This contribution proves that combination of different elements of EE1 proposals works reasonably well and unified filter architecture design is possible. Both multi-scale feature extraction and decomposed convolutions are suggested for unified filter design by multiple proponents during the BoG discussion.