The benchmarks/ layer holds
paper-reproducible experiment code: Open4D methods, reimplemented baselines, configs, metrics, and
scripts — so results can be compared across papers instead of re-derived for each one.
Evaluation in Open4D is systems-aware: a codec is only as good as the pipeline it enables.
Point-to-point and point-to-plane distortion, mesh distances, and perceptual quality of reconstructed geometry versus the source sequence.
Bitrate and compression ratio across operating points — the deciding constraint for streaming 4D content over real networks.
Encode and decode time per frame. Real-time XR playback requires the full decode–render loop to fit inside the display's frame budget.
Consistency of geometry across frames. Per-frame error metrics miss flicker and popping — temporal metrics catch what users actually see.
Every benchmark is explicit and scriptable — method code, baselines, configs, and metrics in one place.
benchmarks/baselines, run under the same harnessopen4d.metrics, identical for methods and baselines# fetch code + baselines
git clone --recurse-submodules \
https://github.com/open4dfoundation/Open4D.git
cd Open4D && pip install -e .
# see available benchmarks and baselines
ls benchmarks/ benchmarks/baselines/
# reproducible environment (optional)
docker build -t open4d docker/Benchmarks are typically run on standard dynamic-geometry datasets used across the field.
| Dataset | Content | Typical use |
|---|---|---|
| 8i Voxelized Full Bodies (8iVFB) | Dynamic point clouds of human performers (longdress, loot, redandblack, soldier) | Point-cloud compression & streaming |
| MPEG dynamic mesh CTC content | Time-varying mesh sequences from the MPEG common test conditions | Mesh codec evaluation (vs. V-DMC-class methods) |
| Volumetric capture sequences | Lab-captured time-varying meshes with changing connectivity | TVMC, tracking, and editing modules |
| Scene reconstructions | Time-varying scene meshes from robotics / autonomy capture | TSMC scene-level compression |
Published benchmark numbers will appear here as modules and papers are released. Until then, the benchmarks directory is the source of truth for what is currently runnable.
Reimplemented baselines are among the most valuable contributions — every method added makes all future comparisons cheaper and fairer.