---
license: cc-by-nc-sa-4.0
task_categories:
- image-segmentation
tags:
- lidar
- semantic-scene-completion
- semantickitti
- autonomous-driving
---

# GSSC-S2D2 — baseline predictions and rare-class object bank

Supporting artefacts for the GSSC / S²D² work on LiDAR semantic scene completion.
This repository holds the **frozen baseline predictions** that S²D² refines, and the
**rare-class object bank** used by the PS³ augmentation pipeline.

> **Looking for the dataset itself?** The paired sparse–dense synthetic scenes are a
> different release: [`Stone-Chern/PS3-SemanticKITTI`](https://huggingface.co/datasets/Stone-Chern/PS3-SemanticKITTI),
> mirroring IEEE DataPort [`10.21227/nqgf-9k39`](https://dx.doi.org/10.21227/nqgf-9k39).
> That is the one to start with.

## Why archives rather than per-frame files

Uncompressed this corpus is 609,349 files and roughly 600 GB, with several directories
at the Hugging Face limit of 10,000 files per directory. The per-frame data compresses
about 110:1, so it ships as ten `zstd` archives totalling 5.4 GB. Every one of the
609,349 original files is present; the member count of each archive is verified against
the source tree.

## Contents

| Archive | Files | Size | What it is |
| --- | ---: | ---: | --- |
| `scpnet_predictions_sequences.tar.zst` | 81,308 | 737 MB | SCPNet predictions, real KITTI sequences 00–21 |
| `scpnet_predictions_synthetic.tar.zst` | 174,063 | 1.14 GB | SCPNet predictions, full synthetic pool |
| `scpnet_predictions_synthetic_31k.tar.zst` | 96,117 | 653 MB | SCPNet predictions, 31K pool |
| `scpnet_predictions_synthetic_30000.tar.zst` | 59,998 | 603 MB | SCPNet predictions, 30K subset |
| `scpnet_predictions_synthetic_10000.tar.zst` | 19,998 | 200 MB | SCPNet predictions, 10K subset |
| `js3cnet_predictions_sequences.tar.zst` | 27,105 | 1.15 GB | JS3C-Net predictions, real sequences 00–21 |
| `js3cnet_predictions_synthetic_filtered.tar.zst` | 38,322 | 162 MB | JS3C-Net predictions, filtered synthetic pool |
| `js3cnet_predictions_synthetic_31k.tar.zst` | 31,442 | 130 MB | JS3C-Net predictions, 31K pool |
| `lmscnet_predictions.tar.zst` | 23,203 | 876 MB | LMSCNet predictions, real sequences |
| `object_bank.tar.zst` | 57,793 | 48 MB | Rare-class object bank (8 classes) used by PS³ |

Each archive carries its own `README.md` and `LICENSE` from the source tree.

## Extracting

```bash
# one archive
tar -I zstd -xf scpnet_predictions_sequences.tar.zst

# or fetch just what you need
huggingface-cli download Stone-Chern/GSSC-S2D2-datasets \
    object_bank.tar.zst --repo-type dataset --local-dir .
```

`zstd` is required (`apt install zstd`, `brew install zstd`, or `pip install zstandard`).

## Licence and attribution

Released under **CC BY-NC-SA 4.0**, inherited from SemanticKITTI, from which this
corpus is derived. Any use must also cite the two papers 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.

The baseline predictions are outputs of SCPNet, JS3C-Net and LMSCNet; please cite those
works when using the corresponding archives.

## Paper

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

## Reference code

[github.com/BillyChern/GSSC-S2D2](https://github.com/BillyChern/GSSC-S2D2)