uvgVPCCenc is an academic open-source volumetric video encoder for the state of the art Video-based Point-Compression (V-PCC) standard. uvgVPCCenc is being developed in C++ under the BSD-3-Clause license.

The uvgVPCC encoder accepts point cloud frames as input (with geometry being positive integers), supports all voxel sizes, and runs on Linux. In version 1.0, the focus is on building a functional encoding pipeline using only essential tools and algorithms inspired by the TMC2 reference software. Prioritizing practical encoding, we’ve omitted lossless tools and those adding significant computational complexity such as geometry reconstruction. In the future we plan on developing innovative methods to reduce the massive complexity of the V-PCC encoding process.

uvgVPCCenc serves as a research platform for new coding tool development and other encoder research activities as well as provides a high-quality and practical V-PCC encoder for the public to use.

uvgVPCCenc is still under development. Speed and RD-quality will continue to improve.

Join our Discord channel to contact us .

https://ultravideo.fi/uvgvpccenc.html for more information.

Compilation and testing

Dependencies

To compile the library and the application, please install CMake and g++. On Debian/Ubuntu:

sudo apt install cmake g++

To generate the documentation, doxygen and graphviz are required. On Debian/Ubuntu:

sudo apt install doxygen graphviz

Compile uvgVPCCenc

To compile the encoder, please use following commands:

cmake --preset=Release

cmake --build --preset=Release

By default, the build application is located in _build/Release/src/app/. Refer to "Using uvgVPCCenc" section to learn how to use the library, and attached application.

Compile with V3C RTP support

To enable sending over RTP run cmake with

cmake --preset=Release -DENABLE_V3CRTP=ON

Test uvgVPCCenc

To test the encoder, please use following commands:

ctest --preset=Release

This runs small tests and checks whether the resulting encoding behaves as expected by verifying the bitstream MD5.

Compile documentation

To generate the doxygen documentation, please use the following:

cmake --preset=docs

cmake --build --preset=docs

The doxygen is then accessible under docs/html/index.html.

Using uvgVPCCenc

To demonstrate how to use the uvgVPCCenc library, a straightforward example application is provided in src/app, with its main() function located in uvgVPCCencAppExample.cpp.

Example:

uvgVPCCenc -i <path_to_ply> -n 10 -o out.vpcc

The mandatory parameters are input, output and the number of frames. The input path should use %0xd for frame numbering. Example sequences are available in the UVG-VPC dataset. Example of an input sequence name: ReadyForWinter_UVG_vox10_25_0_250_%04d.ply.

If the voxel size (aka voxel resolution or geometry precision) is not in the filename, the voxel size must be given: --geo-precision=10.

If the start frame (the index of the first frame of the sequence) is not in the filename, the start frame must be given: --start-frame=0.

uvgVPCCenc accepts .ply files, using positive integer for geometry.

The application option --uvgvpcc accept a string containing uvgVPCCenc parameters, separated by commas. Here is an example command:

uvgVPCCenc -i <path_to_ply> -n 10 -o out.vpcc -t 20 --uvgvpcc rate=16-22-2,presetName=slow,mode=AI

The list of uvgVPCCenc parameters that can be modified with the --uvgvpcc option is defined in the parameterMap initialized in parameters.cpp.

V3C RTP Example:

uvgVPCCenc -i <path_to_ply> -n 10 --dst-address=127.0.0.1 --dst-port=8890

Send encoded bitstream over RTP. Destination can be set using --dst-address and --dst-port. A separate RTP media stream is created for each type of V3C unit and the streames are multiplexed to the same port using SSRC (hardcoded to be vuh_unit_type + 1).

Application parameters

-i, --input <file>           Input filename (using %0Xd)
-o, --output <file>          Output filename
-n, --frames <number>        Number of frames to encode
-s, --start-frame <number>   Frame number to start the encoding
-g, --geo-precision <number> Geometry precision for encoding
-t, --threads <number>       Maximum number of threads to be used (0 auto detection)
-l, --loop-input <number>    Number of input loop (0 for inifinite loop)
    --uvgvpcc <params>       Encoder configuration parameters (see next section)
    --help                   Show this help message
    --version                Show version information
    --dst-address <IP>       Destination IP address for an rtp stream (when compiled with V3C RTP support)";
    --dst-port <number>      Destination port for an rtp stream (when compiled with V3C RTP support)";

uvgVPCCenc parameters

The uvgVPCCenc library includes several options that can only be configured using the API function uvgvpcc_enc::API::setParameter(...). The example application provides the –uvgvpcc <string> option, which passes the specified parameter string to the setParameter API function.

All user-accessible parameters for uvgVPCCenc are defined in the parameterMap variable, located in the file parameters.cpp.

Here are a selection of common uvgVPCCenc parameters:

presetName          Preset name, i.g. slow or fast
rate                V-PCC rate, i.g. 16-22-2
mode                Encoding mode, i.g. AI or RA
logLevel            Level of logging, i.g. TRACE

Presets

Six presets for Voxel 9, 10, and 11, labeled Fast and Slow, have been created to clarify the encoding pipeline parametrization of uvgVPCCenc. These presets were manually crafted starting from Kvazaar's own presets and refined through extensive tool space exploration using stochastic methods. While they are not exhaustive and do not address certain features, such as single- and double-layer options, they provide a practical reference for parameter adjustments. Notably, such presets are exclusive to uvgVPCCenc and do not exist for TMC2.

Best practices for tuning uvgVPCCenc

Volumetric video compression involves a trade-off between speed, quality, and bitrate. uvgVPCCenc provides flexible parameters to match your needs. While presetName=fast and presetName=slow offer general presets, they are not always optimal. Below are tuning tips depending on your goal.

Achieving High Encoding Quality

For the best visual quality, at the expense of slower encoding and higher bitrate:

Set presetName=slow.
Use low QP values, e.g., rate=16-22-2 (down to rate=1-1-2).
Enable lossless video encoding:
- geometryEncodingIsLossless=true
- attributeEncodingIsLossless=true

Notice that full V-PCC lossless is not supported (some 3D points will still be lost due to 2D projection).

Achieving High Encoding Speed (High FPS)

If you want to maximize encoding speed and can trade off some compression efficiency:

Set presetName=fast.
Use high QP values, e.g., rate=32-42-4 (up to rate=51-51-4).
Disable double layer: doubleLayer=false.
Lower internal geometry bit depth:
- geoBitDepthVoxelized=8
- geoBitDepthRefineSegmentation=7

Achieving Low Bitrate

To reduce the bitrate while keeping quality and speed acceptable:

Set presetName=slow.
Use high QP values, e.g., rate=32-42-4.
Reduce the number of encoded points:
- Disable double layer: doubleLayer=false

uvgVPCCenc Version 1.1

Version 1.1 shows better performance with negligible quality degradation or change in compression ratio.

Compared to 1.0, uvgVPCCenc 1.1 features a better implementation of the point cloud and map processing parts (with no impact on 2D encoding speed), resulting in a speed-up of up to 1.7× in Fast voxel 9. Memory management is also significantly improved: uvgVPCCenc 1.1 runs smoothly with a small and stable RAM footprint over time. New developer features and bug fixes are also included in this version. As shown in the results below, compression efficiency (in terms of quality and bitrate) is not affected when switching from 1.0 to 1.1.

The following results were obtained on the 12 sequences of the UVG-VPC dataset, each comprising 64 frames, using all 5 CTC bitrates. Experiments were run on an Intel i7-12700 CPU with 20 threads.

Encoding speed comparison: Major performance boosts in non-video encoding tasks

Voxel Size	Preset	Mode	Version	Speedup	FPS
Voxel 10	slow	RA	V1.0	-	0.8
			V1.1	x1.1	0.9
		AI	V1.0	-	1.0
			V1.1	x1.14	1.2
	fast	RA	V1.0	-	4.7
			V1.1	x1.70	7.9
		AI	V1.0	-	5.0
			V1.1	x1.81	9.1
Voxel 9	slow	RA	V1.0	-	3.7
			V1.1	x1.09	4.0
		AI	V1.0	-	4.4
			V1.1	x1.13	4.9
	fast	RA	V1.0	-	22.3
			V1.1	x1.63	36.3
		AI	V1.0	-	24.8
			V1.1	x1.75	43.4

Compression efficiency comparison: Negligible impact on quality and bitrate

BD-BR PCQM for V1.1 using V1.0 as anchor (negative BD-BR shows better compression efficiency).

Voxel Size	Preset	Mode	BD-BR PCQM
Voxel 10	slow	RA	-1.09%
		AI	0.06%
	fast	RA	-0.89%
		AI	0.01%
Voxel 9	slow	RA	-3.33%
		AI	-0.27%
	fast	RA	-2.30%
		AI	0.12%
		Total average	-0.96%

Academic research

If you use uvgVPCCenc in your research, please cite the most appropriate papers in the following papers:

[1] L. Fréneau, G. Gautier, A. Mercat, and J. Vanne, “uvgVPCCenc: Practical open-source encoder for fast V-PCC compression,” Accepted to ACM Multimedia System, Stellenbosch, South Africa, Mar.—Apr. 2025.

[2] L. Fréneau, G. Gautier, H. Tampio, A. Mercat, and J. Vanne, “Real-Time Video-based Point Cloud Compression,” in Proc. IEEE Visual Comm. Image Proc., Tokyo, Japan, Dec. 2024.

Table of Contents