JVET-W0122 EE2-related: On spatial MV propagation and neighboring template block access for template matching and multi-pass DMVR [C.-C. Chen, C.-T. Hsieh, H. Huang, V. Seregin, W.-J. Chien, Y.-J. Chang, Z. Zhang, Y. Zhang, M. Karczewicz (Qualcomm)]
This contribution proposed a bounding-box approach to limit the number of reference samples fetched from reference pictures within a deterministic size and a replacement method for current block templates by using their corresponding prediction samples through inferred motion vectors for template matching (TM) and multi-pass decoder-side motion vector refinement (MPDMVR) in ECM-1.0. These two were tested on top of ECM-1.0 with the common test condition (JVET-V2024).
Experimental results of Class B sequences with random access test condition are summarized, as follows.
- Test 1 (Bounding-box approach) : 0.99% (Y), 0.99% (U), 1.06% (V), 89% (EncT), 82%(DecT)
- Test 2 (Test 1+Template replacement): 2.04% (Y), 2.20% (U), 2.30% (V), 90% (EncT), 79% (DecT)
- Test 3 (TM off+MPDMVR off) : 3.97% (Y), 4.01% (U), 4.25% (V), 80% (EncT), 60% (DecT)
It was asked if the memory bandwidth might be increased by the method due to the prefetch. The proponents answer that this would not be the case, as the area has the same size as in current DMVR.
It is further discussed that the benefit may be hardware dependent.
As a benefit, it is claimed that it not necessary to rely on motion compensation of the neighbor CU, only motion derivation of the neighbor is needed and causes less latency.
Some of the loss (0.2-0.3%) comes by disabling TMMVP.
Has it been considered using unrefined MVs? No.
At the current stage of exploration, exploring the potential of additional compression gain is still the highest importance, rather than re-designing tools for practicality of implementation. The loss incurred by the given proposal is not negligible, even in the simplest version “Test 1” which does not seem to resolve the problems of DMVR entirely.
Further study was recommended – it would be premature for an EE.