Draco attributes: types, prediction schemes and quantisation

The attribute section holds the values at the vertices: positions, normals, colours, texture coordinates and any other data. Draco compresses each attribute in three steps: it quantises floating-point values to integers, predicts each value from values already decoded, and entropy-codes the small corrections. A decoder reverses them. Attributes are grouped into attribute decoders; with Edgebreaker, each decoder after the first has its own seams in the connectivity.

Attribute decoders

Field	Type	Meaning
`num_attributes_decoders`	UI8	Number of attribute decoders
With Edgebreaker only, for each decoder:
`att_dec_data_id`	UI8	The decoder’s id (“attribute decoder ids” in the specification)
`att_dec_decoder_type`	UI8	0 `MESH_VERTEX_ATTRIBUTE`, one value per vertex; 1 `MESH_CORNER_ATTRIBUTE`, values per corner, split at seams
`att_dec_traversal_method`	UI8	Order in which values are visited: 0 `MESH_TRAVERSAL_DEPTH_FIRST`, 1 `MESH_TRAVERSAL_PREDICTION_DEGREE` (vertices with more decoded neighbours first)
Then, for each decoder:
`att_dec_num_attributes`	varUI32	Number of attributes in this decoder
For each of its attributes:
`att_dec_att_type`	UI8	What the attribute is; see below
`att_dec_data_type`	UI8	The type of each component; see below
`att_dec_num_components`	UI8	Components per value, for example 3 for a position
`att_dec_normalized`	UI8	1 when integer values stand for numbers from 0 to 1 (or −1 to 1)
`att_dec_unique_id`	varUI32	The attribute’s id, which metadata refers to
Then, for each of its attributes:
`seq_att_dec_decoder_type`	UI8	0 generic, 1 integer, 2 quantisation, 3 normals; see below

Attribute type	Name	Data type	Name
0	`POSITION`	1	`DT_INT8`
1	`NORMAL`	2	`DT_UINT8`
2	`COLOR`	3	`DT_INT16`
3	`TEX_COORD`	4	`DT_UINT16`
4	`GENERIC`	5	`DT_INT32`
		6	`DT_UINT32`
		7	`DT_INT64`
		8	`DT_UINT64`
		9	`DT_FLOAT32`
		10	`DT_FLOAT64`
		11	`DT_BOOL`

These two lists come from the Draco library’s source, since the specification names the fields without listing their values. Data type 0 (DT_INVALID) is not a legal value. Builds of the library with transcoder support add tangent, material, joints and weights attribute types after GENERIC, which the source notes would be a change to the bitstream.

Sequential decoder types

Value	Name	Values stored as	Extra data at the end
0	`SEQUENTIAL_ATTRIBUTE_ENCODER_GENERIC`	The attribute’s own data type, unpredicted	None
1	`SEQUENTIAL_ATTRIBUTE_ENCODER_INTEGER`	Predicted integers	None
2	`SEQUENTIAL_ATTRIBUTE_ENCODER_QUANTIZATION`	Floats quantised to integers, then predicted	Quantisation data
3	`SEQUENTIAL_ATTRIBUTE_ENCODER_NORMALS`	Unit normals as two octahedral coordinates, then predicted	Quantisation bits

Attribute values

After the decoder headers, and after the decoder has worked out the order in which to visit the values from the connectivity, each decoder’s attributes follow in turn:

Field	Type	Meaning
`seq_att_dec_prediction_scheme`	I8	The prediction scheme; see below
Unless the scheme is −2 (none):
`seq_att_dec_prediction_transform_type`	I8	How a prediction and a correction combine; see below
`seq_int_att_dec_compressed`	UI8	1 when the corrections are entropy-coded
Corrections	Symbols	`values` × `components` symbols
Prediction data	See below	Data the scheme needs, if any
Transform data	See below	Data the transform needs

Each symbol becomes a signed correction: an even symbol s is s ÷ 2, and an odd one is −(s − 1) ÷ 2 − 1. So 0, 1, 2, 3, 4 stand for 0, −1, 1, −2, 2. After all attributes come the quantisation data, attribute by attribute. The specification covers only predicted, entropy-coded values; attributes with no prediction or with uncompressed corrections are defined by the library’s source alone.

Prediction schemes

Value	Name	Predicts each value from	Prediction data in the file
−2	`PREDICTION_NONE`	Nothing	None
−1	`PREDICTION_UNDEFINED`	Not a stored value	None
0	`PREDICTION_DIFFERENCE`	The previous value in decoding order (the first from 0)	None
1	`MESH_PREDICTION_PARALLELOGRAM`	A parallelogram completed from the opposite triangle: `next` + `previous` − `opposite`	None
2	`MESH_PREDICTION_MULTI_PARALLELOGRAM`	Several parallelograms averaged; in the library’s source, not in the specification	Not described
3	`MESH_PREDICTION_TEX_COORDS_DEPRECATED`	An older texture coordinate scheme; in the library’s source, not in the specification	Not described
4	`MESH_PREDICTION_CONSTRAINED_MULTI_PARALLELOGRAM`	Up to four parallelograms (`kMaxNumParallelograms`), leaving out those across crease edges	Crease flags
5	`MESH_PREDICTION_TEX_COORDS_PORTABLE`	The triangle’s positions and the texture coordinates already decoded, projecting the new point into texture space	Orientations
6	`MESH_PREDICTION_GEOMETRIC_NORMAL`	The normal computed from the surrounding triangles’ positions	Flip bits

The mesh schemes need positions already decoded, so positions are in the first attribute decoder and the other attributes come after.

Scheme	Prediction data
4	For each of the 4 parallelogram counts: `constrained_multi_num_flags` (varUI32); when above 0, a probability (UI8), a size (varUI32) and that many bytes of rABS data giving one crease bit per flag
5	`tex_coords_num_orientations` (UI32), then a probability (UI8), a size (varUI32) and rABS data: one bit per orientation, where 0 flips the previous orientation and 1 keeps it, starting from “true”
6	The transform data first, then a probability (UI8), a size (varUI32) and rABS data: one bit per value saying whether to flip the predicted normal

Prediction transforms

Value	Name	Data	Meaning
−1	`PREDICTION_TRANSFORM_NONE`	None	Not a stored value
0	`PREDICTION_TRANSFORM_DELTA`	None	Value = prediction + correction; in the library’s source
1	`PREDICTION_TRANSFORM_WRAP`	min I32, max I32	The prediction is clamped to min…max; value = clamped prediction + correction, wrapped back into min…max by adding or subtracting (max − min + 1)
2	`PREDICTION_TRANSFORM_NORMAL_OCTAHEDRON`	Not described	Octahedral normal transform; in the library’s source
3	`PREDICTION_TRANSFORM_NORMAL_OCTAHEDRON_CANONICALIZED`	max_q_val I32, then an unused I32	Corrections to octahedral normal coordinates, worked in a canonical quarter of the octahedron so that they stay small

With transform 3, the symbols are taken as they are, without the signed conversion above.

Quantisation

For decoder type 2, after the values:

Field	Type	Meaning
`quantized_data_min_values`	Float × components	The smallest value of each component
`quantized_data_max_value_df`	Float	The range: the largest difference from the minimum over all components
`quantized_data_quantization_bits`	UI8	Bits per quantised value

A quantised integer q becomes min[c] + q × range ÷ (2^bits − 1), with the sign of q kept. For decoder type 3 only the quantisation bits (UI8) are stored: the two integer octahedral coordinates, each up to 2^bits − 1, are turned back into a unit vector.