JVET-AF0161 AHG12: Harmonized IBC/ITMP search area with adaptive sampling [D. Ruiz Coll, B. Chen (Ofinno), K. Naser, P. Bordes, F. Le Léannec, F. Galpin (InterDigital)]

This contribution proposes a harmonization of the IBC and IntraTMP search areas. The IBC and IntraTMP tools use the reconstructed samples of the current picture to select a candidate prediction block for the current block. The IBC search area comprises a constant number of CTUs, while the IntraTMP search area is based on the current block dimension, which is mainly smaller than the IBC search area. IntraTMP searches for the most similar L-shaped, Top-Only, or Left-Only template that matches the template of the current block in the search area. In IntraTMP, a constant sampling factor of 3 is applied in the horizontal and vertical search directions to alleviate the template-matching complexity. However, the extension of the IntraTMP search area demands more searches. This contribution proposes two methods to determine an adaptive sampling factor when the IntraTMP search region is extended to the IBC search region.

On top of ECM-10.0, simulation results are reported as follows:

Test-1: Sampling rate based on the searching distance.

AI: Class F: {-0.43%, -0.35%, -0.28%, 100.9%, 105.0%},

AI: Class TGM: {-0.70%, -0.72%, -0.73%, 101.5%, 108.3%};

RA: Class F: {-0.16%, -0.00%, -0.04%, 100.5%, 100.5%},

RA Class TGM: {-0.43%, -0.54%, -0.52%, 100.2%, 102.6%};

Test-2: Sampling rate based on the IntraTMP regions dimension.

AI: Class F: {-0.29%, -0.26%, -0.14%, 101.4%, 104.1%},

AI: Class TGM: {-0.43%, -0.43%, -0.48%, 100.4%, 101.1%};

Test-3: Combination of Test-1 and Test-2.

AI: Class F*: {-0.32%, -0.32%, -0.24%, 100.8%, 101.5%},

AI: Class TGM: {-0.32%, -0.33%, -0.36%, 101.1%, 100.0%};

Contribution JVET-AF0115 also extends IBC search range.

It was asked if this was tested on camera-view content. An earlier interation of the proposal was tested and showed unfavourable performance vs. runtime tradeoff for camera-view content. For this current proposal, it could be tested again on camera-view content to see the performance vs. runtime tradeoff.

This proposal changes the search range of IntraTMP, and uses the entire area of IBC search range. It further applies two elements to the IntraTMP search area: 1) sampling rate based on search distance, and 2) sampling rate based on dimensions.

Since JVET-AF0115 extends IBC search range, and the concept of this proposal is to use the IBC search range, then thre is camera-view interaction between the two proposals.

It was commented that combining the two elements in this proposal would actually result in worse performance (Test 3 vs. Test 1 and Test 2). However, the proponents said that Test 3 was mainly targeted at reducing decoding time.

It was agreed to investigate this in an EE, to test the two elements separately and in combination, and also to test the interaction between this proposal and JVET-AF0115 in the form of additional combination tests. Results on camera-view content were also requested.