.HEVC File Format – 3D-HEVC

Introduction

Annex I of ITU-T H.265 (01/2026, version 11) defines 3D-HEVC. It is built on the multi-layer structure of Annex F: each view has a texture layer and a depth layer (DepthLayerFlag 1, dimension 0 of the scalability mask), and the coding tools predict between views and between a texture layer and its depth layer. The syntax of Annexes F and G applies except where this page replaces it. In the 3D Main profile depth layers are monochrome (4:0:0). The tools it adds are: disparity vectors derived from depth, inter-view merging and residual prediction, view synthesis prediction, illumination compensation, depth-based block partitioning (DBBP), and, for depth layers, wedgelet and contour intra modes, DC-only blocks and prediction of the coding quadtree from the texture. The 3D Main profile is general_profile_idc 8.

VPS 3D extension (I.7.3.2.1.7)

Camera parameter conversion from depth to disparity. In Annex I the VPS has, after vps_extension(), vps_extension2_flag; when it is 1 then vps_3d_extension_flag u(1), and when that is 1 alignment bits equal to 1 and vps_3d_extension(), then vps_extension3_flag and extension data bits. ViewOIdxList[n] lists the view order indices of the views of the VPS.

SPS 3D extension (I.7.3.2.2.5)

Present when sps_3d_extension_flag is 1 (SPS and the multi-layer SPS). Two copies of each flag are sent, d = 0 for texture layers and 1 for depth layers (DepthFlag). Every flag is inferred 0 when absent.

PPS 3D extension and depth look-up tables (I.7.3.2.3.7, I.7.3.2.3.8)

A depth look-up table (DLT) maps the few depth values that actually occur to indices, which makes depth DC offsets cheaper to code. It lives in the PPS.

The table of a layer from delta_dlt() holds every value from 0 to depthMaxValue for which "is in the reference table" (when predicted, else false) differs from "is in the delta list" (I-16): the list toggles entries of the prediction.

Slice segment header (I.7.3.6.1)

The multi-layer header with these differences and additions. Annex I uses NumRefListLayers where Annex F has NumDirectRefLayers in the inter-layer conditions.

Coding quadtree and coding unit (I.7.3.8.4, I.7.3.8.5)

Annex I replaces the coding quadtree and coding unit syntax; the rest of slice segment data (prediction unit, transform tree, residual, palette, …) is as in Annex F. The variables CqtCuPartPredEnabledFlag, IntraContourEnabledFlag and the other capitalised flags are the SPS flags of the layer's DepthFlag (set up in the slice semantics together with DispAvailFlag, RpRefPicAvailFlag and DisparityDerivationFlag).

Changes to the ordinary elements

partPredIdc, for depth layers with CqtCuPartPredEnabledFlag: 0 when the texture CU has a different size or is 2N×2N or N×N; 1 when it is 2N×N, 2N×nU or "2NxnL" (as printed in the text; no such partition exists, so 2N×nD is probably meant); otherwise 2. The part mode table (Table I.1) is the ordinary one for partPredIdc 0 (intra 2N×2N and N×N; inter 0 to 7 as 2N×2N, 2N×N, N×2N, N×N, 2N×nU, 2N×nD, nL×2N, nR×2N), and for partPredIdc 1: 0 2N×2N, 1 2N×N, 2 2N×nU, 3 2N×nD; for 2: 0 2N×2N, 1 N×2N, 2 nL×2N, 3 nR×2N. Its binarisation (Table I.16) lists the ordinary codes for partPredIdc 0 and gives the restricted forms short codes: partPredIdc 1 gives 2N×N 0 (01 with AMP above the minimum size), 2N×nU 000 and 2N×nD 001; partPredIdc 2 gives N×2N 0 (01), nL×2N 000 and nR×2N 001; 2N×2N is 1 in every case.

Added elements

A CU codes cu_extension() after its prediction data, then depth_dcs() when it is DC-only or intra and not skip_intra, then the transform tree unless it is skipped, skip_intra, DC-only or PCM. Intra prediction units loop over the CU (or its four parts for N×N) sending intra_mode_ext() first for each, then the usual MPM fields only where no_dim_flag is 1. A DC-only or DIM block predicts each of its one or two segments from the DLT-mapped DC offset; the intra sample processes (INTRA_WEDGE, INTRA_CONTOUR, INTRA_SINGLE, segment DC derivation) are in I.8.4.

CABAC for the added elements (I.9.3)

Appended to the tables of the HEVC CABAC page (Tables I.4 to I.18). initType is 0, 1 or 2 as there; the tables below give initValue per initType.

The CU part mode binarisation changes as described above (Table I.16). All other elements use the single-layer rules.

3D Main profile (I.11.1.1)

Indicated by general_profile_idc 8 (or compatibility flag 8) with general_max_12bit_constraint_flag, general_max_10bit_constraint_flag, general_max_8bit_constraint_flag, general_max_422chroma_constraint_flag, general_max_420chroma_constraint_flag and general_lower_bit_rate_constraint_flag 1 and the monochrome, intra and one-picture flags 0 (the sub_layer_ forms for a lower temporal operation point). The base layer is Main. Texture layers are 4:2:0 and depth layers 4:0:0, all 8-bit. The VPS has at most 16 representation formats. Only the view order and depth dimensions are used. Alternative output layers are off. Layers other than the first have no reference location offsets, the same picture size as their reference layers, and no colour mapping. No range extension or SCC syntax, CtbLog2SizeY 4 to 6, a layer has at most 9 direct and indirect reference layers, tiles are at least 256 × 64 and not combined with wavefronts. Layers of one kind have increasing view order index, and a depth layer directly follows the texture layer of its view in the VPS. Level limits (I.11.2) are not reproduced.

SEI messages

The messages of Annexes F and G apply, including the multiview and depth ones. Annex I adds type 181.

Not covered on this page

The 3D decoding processes of I.8 (candidate lists for disparity vectors, DLT derivation, depth intra prediction, merge candidates, illumination compensation, bilinear residual prediction, DBBP and view synthesis), the wedgelet pattern table generation (I.6.6), the 3D level limits (I.11.2) and the 3D forms of the scalable nesting SEI semantics. Source: ITU-T H.265 (01/2026, version 11), Annex I.