README.md

---
license: cc-by-nc-sa-4.0
pretty_name: PS3-SemanticKITTI
annotations_creators:
  - machine-generated
task_categories:
  - image-segmentation
size_categories:
  - 10K<n<100K
tags:
  - semantic-scene-completion
  - lidar
  - point-cloud
  - semantickitti
  - autonomous-driving
  - synthetic
  - 3d
viewer: false
---

# PS3-SemanticKITTI

Paired sparse-dense synthetic scenes for LiDAR semantic scene completion.

Every dense semantic scene in this corpus is paired with a matched sparse LiDAR
observation of the same scene, so a model can be trained on the sparse-to-dense
mapping without either half having been recorded by a real sensor.

## What is in the deposit

| Archive | Scenes | Compressed |
| --- | --- | --- |
| `synthetic_pool_31K.tar.gz` | 32,039 | ~2.2 GiB |
| `synthetic_pool_57K.tar.gz` | 57,650 | ~3.9 GiB |

Both archives are also mirrored on Hugging Face, byte for byte identical to
the copies deposited here and downloadable without an IEEE DataPort
subscription:

<https://huggingface.co/datasets/Stone-Chern/PS3-SemanticKITTI>

The mirror becomes public alongside the reference code linked at the end of
this file. Cite the DOI whichever copy you download: `10.21227/nqgf-9k39` is
this dataset's citable identifier of record, and the mirror carries the same
`LICENSE`.

**Take the 31K pool unless you specifically want the larger variant.** It is the
pool the accompanying work trains on for its reported results.

`synthetic_pool_31K` is a **strict subset** of `synthetic_pool_57K`, verified over
all 96,117 of its files, each resolving to the same content. Downloading both
duplicates 32,039 scenes.

The accompanying work reports the larger pool as a **negative result**: scaling
beyond 32,039 scenes did not improve accuracy. The 57K archive is released so
that finding can be checked, not because more data is better here.

## Layout

Each archive extracts to a flat directory holding three NumPy arrays per scene.

```text
<id>_voxels.npy     (256, 256, 32)  uint8   sparse binary occupancy
<id>_gt_scene.npy   (256, 256, 32)  uint8   dense semantic labels, 0-19
<id>_bev.npy        (256, 256)      uint8   bird's-eye-view map, 0-19
```

Each archive also contains `LICENSE` and `generation_metadata.json`.

`_voxels.npy` is strictly binary, with values `{0, 1}`, and marks which voxels
the simulated sensor returned. `_gt_scene.npy` is the complete scene from which
that sweep was traced.

```python
import numpy as np

sweep = np.load("000000_voxels.npy")    # (256, 256, 32) uint8, {0, 1}
dense = np.load("000000_gt_scene.npy")  # (256, 256, 32) uint8, 0..19
bev = np.load("000000_bev.npy")         # (256, 256)     uint8, 0..19
```

## Classes

Label `0` is free or unlabelled space. Labels `1-19` follow the SemanticKITTI
learning map.

| Label | Class | Label | Class |
| --- | --- | --- | --- |
| 0 | free / unlabelled | 10 | parking |
| 1 | car | 11 | sidewalk |
| 2 | bicycle | 12 | other-ground |
| 3 | motorcycle | 13 | building |
| 4 | truck | 14 | fence |
| 5 | other-vehicle | 15 | vegetation |
| 6 | person | 16 | trunk |
| 7 | bicyclist | 17 | terrain |
| 8 | motorcyclist | 18 | pole |
| 9 | road | 19 | traffic-sign |

This is the standard 20-way learning map, so the labels are directly comparable
with SemanticKITTI annotations and with the output of models trained on them.

## How the scenes were made

1. **Generation.** Three cascaded coarse-to-fine multinomial diffusion models,
   at 32×32×4, then 64×64×8, then 256×256×32, trained on the
   SemanticKITTI training split.
2. **Screening.** Each candidate's class histogram is compared against a corpus
   prior estimated on held-out real scenes, and a scene is kept only where the
   Jensen-Shannon divergence falls below a threshold. This is what stops the
   generator drifting toward implausible class mixtures.
3. **Rare-class insertion.** Surviving scenes receive objects from a rare-class
   object bank, placed on ground-level voxels.
4. **Sensor simulation.** A hardware-aware HDL-64E ray-tracer converts each
   dense scene into the sparse sweep that sensor would have returned. Pasting
   happens before ray-tracing, so the sparse half of each pair genuinely
   contains the inserted rare classes.

The pipeline exists for the long tail. It produces supervision for the scarcest
classes at source, rather than reweighting a distribution that never contained
enough of them.

`generation_metadata.json` inside each archive records that pool's aggregate
provenance. For the 57K pool, per-batch generation records were not retained.
The one surviving batch record covers 2,055 scenes and ships alongside as
`generation_metadata_batch_seed2024.json`; it does not describe the pool, and
the metadata file says so.

## Licence and attribution

Released under **CC BY-NC-SA 4.0**. This is inherited rather than chosen: the
corpus is derived from SemanticKITTI, which carries the same terms, so the
non-commercial and share-alike conditions follow through to this data and to
anything derived from it.

Cite this dataset as:

> Shi Chen, Weifeng Ge, "PS3-SemanticKITTI: Paired Sparse-Dense Synthetic Scenes
> for LiDAR Semantic Scene Completion", IEEE Dataport, August 23, 2026,
> doi:10.21227/nqgf-9k39

```bibtex
@data{nqgf-9k39-26,
  doi       = {10.21227/nqgf-9k39},
  url       = {https://dx.doi.org/10.21227/nqgf-9k39},
  author    = {Shi Chen and Weifeng Ge},
  publisher = {IEEE Dataport},
  title     = {PS3-SemanticKITTI: Paired Sparse-Dense Synthetic Scenes for
               LiDAR Semantic Scene Completion},
  year      = {2026}
}
```

`@data` is IEEE Dataport's own entry type; substitute `@misc` if your
bibliography style does not know it.

Any use must also cite both of the papers that SemanticKITTI requires:

1. J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss and
   J. Gall. *SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR
   Sequences.* ICCV 2019.
2. A. Geiger, P. Lenz and R. Urtasun. *Are we ready for Autonomous Driving? The
   KITTI Vision Benchmark Suite.* CVPR 2012, pp. 3354-3361.

## Paper

Generative Semantic Scene Completion — https://arxiv.org/abs/2608.26737

## Reference code

The loader, the training configurations these pools were built for, and the
evaluation protocol are at
[github.com/BillyChern/GSSC-S2D2](https://github.com/BillyChern/GSSC-S2D2).