One Hat Cyber Team

View File Name : Appendix-Subpel-Interpolation-Open-Loop-ME.md

# Subpel Interpolation in the Open-loop Motion Estimation appendix

## 1. Description of the algorithm

The open loop motion estimation (ME) uses source input pictures as
reference pictures and provides motion vectors with up to quarter-pel
precision. The luminance samples at the half-pel positions are generated
first by applying symmetric 4-tap filter F1 = {-4, 36, 36, -4} on the
integer luminance samples. The values of the half-pel samples are
generated for the whole input search region and stored in 3 different
buffers. Then, the luminance samples at the quarter-pel positions are
generated using a bilinear filter F2 = {1,1} that is applied on the
full-pel luminance samples and the already computed half-pel luminance
samples.

In Figure 1 below, A00, A10, A01 and A11 represent the luminance samples
at the integer positions; b, h and j are the luminance samples at the
half-pel positions and a, c, d, e, f, g, i, k, n, p, q and r are the
luminance samples at the quarter-pel positions.

![sioplm_fig1](./img/sioplm_fig1.png)

##### Figure 1. Fractional positions.

Table 1 below summarizes the filters used to generate the luminance
samples at the half-pel positions.

##### Table 1. Interpolation filters used in the generation of the half-pel positions in ME.

| **Position** | **Frac-pos Y** | **Frac-pos X** | **Horizontal filter** | **Vertical filter** |
| ------------ | -------------- | -------------- | --------------------- | ------------------- |
| **b**        | 0              | 2              | F1                    | \-                  |
| **h**        | 2              | 0              | \-                    | F1                  |
| **j**        | 2              | 2              | F1                    | F1                  |

Table 2 below summarizes the filters used to generate the luminance
samples at the quarter-pel positions.

#####  2. Interpolation filters used in the generation of the quarter-pel positions in ME.

| **Position** | **Frac-pos Y** | **Frac-pos X** | **Filter**   | **Interpolated sample value** |
| -------------| -------------- | -------------- | ------------ | ----------------------------- |
| **a**        | 0              | 1              | F2           | A00 + b                       |
| **c**        | 0              | 3              | F2           | b + A10                       |
| **d**        | 1              | 0              | F2           | A00 + h                       |
| **e**        | 1              | 1              | F2           | h + b                         |
| **f**        | 1              | 2              | F2           | b + j                         |
| **g**        | 1              | 3              | F2           | b + h'                        |
| **i**        | 2              | 1              | F2           | h + j                         |
| **k**        | 2              | 3              | F2           | j + h'                        |
| **n**        | 3              | 0              | F2           | h + A01                       |
| **p**        | 3              | 1              | F2           | h + b'                        |
| **q**        | 3              | 2              | F2           | j + b'                        |
| **r**        | 3              | 3              | F2           | h' + b'                       |

## 2.  Implementation of the algorithm

The sub-pel search in motion estimation is invoked after the full-pel
search is completed and the best integer position is determined. The
half-pel/quarter-pel search is a refinement stage around the best selected
position in the previous search stage. This refinement area is
configurable depending on the desired quality/complexity tradeoff. For
example, in Figure 2, a first tier of half-pel positions around the best
full-pel position are checked and then a second and third tier of
half-pel positions can be also checked. Similarly, during the
quarter-pel refinement, various tiers of quarter-pel refinement
positions can be checked. The enabling of the multi-tier refinement is
controlled by the half\_pel\_mode flag for half-pel refinement and by
the quarter\_pel\_mode flag for quarter-pel refinement. Moreover, the number
of tiers to be considered in the refinement is determined by the
H\_PEL\_SEARCH\_WIND and the Q\_PEL\_SEARCH\_WIND macros for half-pel and
quarter-pel refinements, respectively.

![sioplm_fig2](./img/sioplm_fig2.png)

##### Figure 2. Subpel search refinement positions.

The description of the half-pel and quarter-pel modes is provided
in Table 3 and Table 4, respectively. The half-pel and quarter-pel modes
for the different encoder presets are summarized in Table 5.

##### Table 3. half\_pel\_mode settings and description.

| **half\_pel\_mode**     | **Description**                 |
| ----------------------- | ------------------------------- |
| 0 (EX\_HP\_MODE)        | Exhaustive 1/2-pel search mode. |
| 1 (REFINMENT\_HP\_MODE) | Refinement 1/2-pel search mode. |

##### Table 4. quarter\_pel\_mode settings and description.

| **quarter\_pel\_mode**  | **Description**                 |
| ----------------------- | ------------------------------- |
| 0 ( EX\_QP\_MODE)       | Exhaustive 1/4-pel search mode  |
| 1 (REFINMENT\_QP\_MODE) | Refinement 1/4-pel search mode. |

##### Table 5. Half-pel and quarter-pel search modes as a function of the encoder preset.

| **Encoder Preset** | **half\_pel\_mode**     | **quarter\_pel\_mode**  |
| ------------------ | ----------------------- | ----------------------- |
| 0                  | 0 (EX\_HP\_MODE)        | 1 (REFINMENT\_QP\_MODE) |
| 1 to 8             | 1 (REFINMENT\_HP\_MODE) | 1 (REFINMENT\_QP\_MODE) |

## Notes

The feature settings that are described in this document were compiled at v0.8.3 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.