JVET-G0023 Qualcomm’s response to Joint CfE in 360-degree video category [M. Coban, G. Van der Auwera, M. Karczewicz (Qualcomm)]

Discussed Thursday 15:00 (GJS & JRO).

The interior of the encoding/decoding processes is not modified. This contribution is only about a projection mapping.

An adjusted cubemap projection (ACP) with padding is submitted to JVET’s Call for Evidence on 360-degree video coding. ACP was previously described in F0025 but without padding. The padded ACP reportedly offers efficient compression of 360-degree video and reduced seam artefacts in the rendered viewports. ACP is an enhancement of CMP by adjusting the sampling on the cube faces to be nearly uniform. With the proposed padding scheme for ACP, the RA coding gain is reportedly 10.7% (e2e WS-PSNR) compared with the equirectangular projection in accordance with the JVET common test conditions for 360-degree video, while for CfE conditions, the reported coding gain is reportedly 8.1% (e2e WS-PSNR).

The scheme is included in 360Lib version 3.

F0025 had slightly higher objective coding gain.

A 3x2 layout of the cubemap is used, with “padding” around 3x1 regions. The “padding” is done by extending the rectangular faces, not by duplicating samples from spatially adjacent faces.

To compensate for the increased number of samples used for the “padding” without increasing the total coded picture size, the number of “active” samples is reduced.

The “padding” was 4 luma samples wide.

Blending is not used for the duplicated regions. The “padding” areas are discarded after decoding.

It was remarked that padding should mainly be useful with regard to subjective boundary artefacts. In discussions it was noted that an alternative to padding is geometry-aware operation of motion compensation could be an alternative to padding.