The Motion Estimation (ME) process generates inter-prediction candidates using highly parallelizable, open loop, neighbor-independent methods. In the current SVT-AV1 encoder, the ME process is based on the input pictures, i.e. the reference pictures are replaced by the corresponding source pictures. As a result, the ME is an open loop operation. The Motion Estimation (ME) process has access to the current input picture as well as to the input pictures the current picture uses as references according to the hierarchical prediction structure under consideration. The ME process is multithreaded, so pictures can be processed out of order as long as corresponding reference input pictures are available. The ME process generates motion estimation information for all square blocks starting at 8x8 size up to the superblock (SB) size. The motion vector information is generated at full-pel precision. The candidates’ costs generated in the ME processes are further refined in downstream processes and as more neighbor information becomes available allowing for more accurate costs to be calculated.
ME process involves four components: Pre-Hierarchical Motion Estimation (pre-HME), Hierarchical Motion Estimation (HME), Search Center Selection, and Motion Estimation (ME). The pre-HME/HME performs a quick search on down-sampled pictures to converge to a candidate search center for full-pel motion estimation search at the full picture resolution. Search center selection selects a single search center from several pre-HME/HME candidates and external search center candidates (in the current configuration (0,0) is always considered as a search centre candidate). One goal of this design is to eliminate the need for large search areas with search center selection and then to fully search a smaller search area with the refinement stage. Motion Estimation finds the best motion vector around the SB search center for each of the partitions being considered.
The goal of pre-HME is to catch very high motion along the horizontal or vertical directions. A narrow, long search is performed along each direction on the sixteenth-downsampled pictures (i.e. downsampling by a factor of 4 in each direction) as shown in Figure 1.
The best overall pre-HME search centre/SAD pair is saved from the search. After HME-level-0, if the best pre-HME result is better than any of the HME-level-0 search centres (based on SAD), the worst HME-level-0 search centre is replaced by the best pre-HME result (to be considered in the next HME stages). As such, HME-level-1 must be enabled to use pre-HME.
Figure 1. Example of a pre-HME search area around point X. The search region is long and narrow in the horizontal and vertical direction.
Hierarchical Motion Estimation (HME) takes as input an enhanced input picture and reference picture and produces a search center for each SB, to be searched at ME. The enhanced input picture is the temporally filtered source picture. The HME consists of up to three stages: a one-sixteenth resolution Level-0 full search, a one-quarter resolution Level-1 refinement search, and a base-resolution Level-2 refinement search as depicted in Figure 2. In addition, the total search area is subdivided into N-search areas, where each of the Level-0, Level-1, and Level-2 searches are performed independently to produce N-search centers. Of the N-search centers, one search center is finally selected. Having multiple search centers prevents the Level-0 and Level-1 searches from choosing local minima and missing the true center of motion completely. Currently, N is 4 for all presets (the search region at HME-level-0 is divided into a 2x2 grid). All search and selection decisions are based on a pure SAD distortion metric. Figure 3 depicts an example HME full search and refinement data flow through Level-0, Level-1, and Level-2.
Search Center Selection chooses the best SB search center for Motion Estimation based on a SAD distortion metric. Search center candidates may be generated from HME or other sources of candidate search centers. A diagram showing search center selection and the Motion Estimation is given in Figure 4.
Figure 4: Search centre selection and motion estimation. Reference N search centre candidates come from HME (or other sources, if applicable).
Motion Estimation (ME) takes as input an enhanced input picture, reference picture, and search center for each SB. ME produces Motion Vectors (MVs), one for each of the 8x8 and larger square blocks in an SB. ME is an integer full search around the search centre on the full resolution picture and is performed for square blocks only. The integer full search produces an integer MV candidate for each 8x8 and larger square blocks and the SB SAD estimation using the base 8x8 block SAD data. As shown in Figure 5, the ME search is performed on 8x8 blocks, and the MV/SAD information for larger block sizes are derived from the 8x8 results (by adding together all the 8x8 SADs that make up a given block).
Figure 5: ME search for the case of a 64x64 SB. The SAD is computed for each 8x8 block. For larger square blocks, the 8x8 SADs are summed to produce the output SAD of the larger blocks.
Inputs: Source pictures: the current frame and all its references in one-sixteenth resolution, one-quarter resolution, and base resolution
Outputs: MVs for each 8x8 and larger square blocks and SB distortion (SAD) values.
The flow of data in open-loop ME is illustrated in Figure 6, along with the relevant functions that are associated with each part of the algorithm.
| Signal | Description |
|---|---|
| skip_search_line | If true, skips every other search region line |
| l1_early_exit | Skip pre-HME search for list 1 references if the SAD from list 0 is low, or the list 0 MVs are small |
| use_tf_motion | Use TF motion to direct prehme searches (if TF motion is horizontal (vertical), search only horizontal (vertical)) |
The search areas of pre-HME, HME and ME can be adjusted for quality/complexity trade-offs. Larger search areas will capture more motion, thereby improving quality, while smaller search areas will require less computation and favour speedups.
For ME and HME, a maximum and a minimum search area are specified. Closer frames are expected to have less motion relative to the current frame, thus the search areas are scaled-up more for distant frames. The actual area searched for each reference frame depends on the distance of the reference frame to the current frame (and will always be greater than or equal to the minimum area and less than or equal to the maximum area). Closer frames are expected to have less motion relative to the current frame, thus the search areas are scaled-up more for distant frames.
Pre-HME level is set with prehme_level, and the search areas are set in set_prehme_ctrls(). ME and HME search areas are set in set_me_hme_params_oq().
The HME-level-0 search region can also be adjusted. HME adjustment is done based on the distance of the reference frames and the results of previously processed
reference frames. Reducing the search area based on distance from the current frame is controlled by distance_based_hme_resizing.
Search region adjustment is also done based on the HME search centre chosen by the first reference frame (list 0, index 0).
Subsequent frames can see whether the motion on the first reference frame was mainly vertical, mainly horizontal, or still.
If the motion is vertical (horizontal), the horizontal (vertical) search area will be reduced. If the motion is still (i.e. low motion) then both dimensions are
reduced. The thresholds for characterising motion as vertical, horizontal, or still are set by reduce_hme_l0_sr_th_min and reduce_hme_l0_sr_th_max.
The ME search region can be adjusted based on the HME results. Specifically, if the HME distortions are low, the block is expected to have low motion, so the ME search region is reduced. Similarly, if the HME output search centre is close to (0,0) the ME search region can be reduced.
IF (HME_search_centre_x <= MV_TH &&
HME_search_centre_y <= MV_TH &&
HME_SAD < stationary_sad_th)
Divide ME search width and height by stationary_divisor
ELSE IF (HME_SAD < general_sad_th)
Divide ME search width and height by general_divisor
Where the variables in the pseudo-code above correspond to the following signals in the code (set in set_me_sr_adjustment_ctrls()):
| Signal in code | Name in pseudo-code | Description |
|---|---|---|
| reduce_me_sr_based_on_mv_length_th | MV_TH | Reduce the ME search region if HME MVs and HME sad are small |
| stationary_hme_sad_abs_th | stationary_sad_th | Reduce the ME search region if HME MVs and HME sad are small |
| stationary_me_sr_divisor | stationary_divisor | Reduction factor for the ME search region if HME MVs and HME sad are small |
| reduce_me_sr_based_on_hme_sad_abs_th | general_sad_th | Reduce the ME search region if HME sad is small |
| me_sr_divisor_for_low_hme_sad | general_divisor | Reduction factor for the ME search region if HME sad is small |
Reference frames can be pruned at early stages of open-loop ME to save the cost of searching them at subsequent, more expensive stages and MD. Pruning is performed on a per-SB basis after HME (before ME) and after ME (so that candidates will not be passed to MD). Pruning decisions are based on the relative SAD of each reference frame (relative to the best SAD), as follows:
Reference pruning controls are set in set_me_hme_ref_prune_ctrls(), with the following controls:
| Signal | Description |
|---|---|
| enable_me_hme_ref_pruning | |
| prune_ref_if_hme_sad_dev_bigger_than_th | TH used to prune references based on HME sad deviation |
| prune_ref_if_me_sad_dev_bigger_than_th | TH used to prune references based on ME sad deviation |
| protect_closest_refs | If true, do not prune closest ref frames |
If the (0,0) SAD is low, pre-HME and HME can be skipped, and the ME search area can be reduced. The feature is controlled by me_early_exit_th,
which is the threshold used to determine if the (0,0) SAD is low enough to apply the optimizations.
The feature settings that are described in this document were compiled at v0.9.0 of the code and may not reflect the current status of the code. The description in this document represents an example showing how features would interact with the SVT architecture. For the most up-to-date settings, it's recommended to review the section of the code implementing this feature.