README.md

---
license: other
license_name: mixed-non-commercial
license_link: https://github.com/YangDongHeon/SpaceCache#license
library_name: pytorch
pipeline_tag: other
tags:
  - point-cloud
  - semantic-segmentation
  - lidar
  - streaming
  - efficient-inference
  - arxiv:2610.21207
---

# SpaceCache checkpoints

Checkpoints for **SpaceCache: Intermediate Feature Reuse for Efficient Semantic
Segmentation on Point Cloud Streams**
([paper](https://arxiv.org/abs/2610.21207),
[code](https://github.com/YangDongHeon/SpaceCache)).

SpaceCache is training-free: these are the backbone segmentation models it
runs on, Utonia and LitePT-B, trained on Waymo, nuScenes and ScanNet.

## Usage

From the root of the [SpaceCache repository](https://github.com/YangDongHeon/SpaceCache):

```bash
huggingface-cli download honeydong49/SpaceCache --include "exp/*" --local-dir .
```

The files land under `exp/`, where the evaluation configs read them.

## Checkpoints

| Backbone | Dataset | File |
|---|---|---|
| Utonia | Waymo | `exp/utonia/semseg-utonia-v1m1-5b-waymo-dec/model/model_best.pth` |
| Utonia | nuScenes | `exp/utonia/semseg-utonia-v1m1-4b-nuscenes-dec/model/model_best.pth` |
| Utonia | ScanNet | `exp/utonia/semseg-utonia-v1m1-0b-scannet-dec/model/model_best.pth` |
| LitePT-B | Waymo | `exp/litept/waymo/semseg-litept-v1m1-0-base/model/model_best.pth` |
| LitePT-B | nuScenes | `exp/litept/nuscenes/semseg-litept-v1m1-0-base/model/model_best.pth` |
| LitePT-B | ScanNet | `exp/litept/scannet/semseg-litept-v1m1-0-base/model/model_best.pth` |

Each file holds the model weights (`state_dict`) only. The Utonia models are
a segmentation decoder trained on top of the released
[Utonia](https://huggingface.co/Pointcept/Utonia) encoder, which was kept
frozen; the file contains both. The LitePT-B models were trained from scratch
with the [LitePT](https://github.com/prs-eth/LitePT) architecture.

## License

All checkpoints are for non-commercial research only.

| Checkpoints | Terms |
|---|---|
| Utonia (all datasets) | contain the [Utonia](https://huggingface.co/Pointcept/Utonia) weights, CC BY-NC 4.0, plus the Waymo agreement (the Utonia encoder was pretrained partly on Waymo; notice below) and the terms of the dataset below |
| Waymo | [Waymo Dataset License Agreement for Non-Commercial Use](https://waymo.com/open/terms) (notice below; copy in `WAYMO_LICENSE.txt`) |
| nuScenes | CC BY-NC-SA 4.0, [nuScenes terms of use](https://www.nuscenes.org/terms-of-use) |
| ScanNet | [ScanNet Terms of Use](http://kaldir.vc.in.tum.de/scannet/ScanNet_TOS.pdf) |

The checkpoints trained on Waymo, and all Utonia checkpoints (the Utonia
encoder was pretrained partly on Waymo), were made using the Waymo Open Dataset,
provided by Waymo LLC under the Waymo Dataset License Agreement for
Non-Commercial Use, available at waymo.com/open/terms, and your access and use
of such work are governed by the terms and conditions therein. They are
released under that agreement; any further downstream use or modification of
them is subject to its terms, including the non-commercial restrictions of its
Section 4. They must not be used to operate or assist the operation of a
vehicle, in production systems, or for other primarily commercial purposes.
A copy of the agreement is in `WAYMO_LICENSE.txt`; the current terms at
waymo.com/open/terms govern.

## Citation

```bibtex
@article{yang2026spacecache,
  title   = {SpaceCache: Intermediate Feature Reuse for Efficient Semantic Segmentation on Point Cloud Streams},
  author  = {Yang, Dongheon and Kang, Beomseok and Kim, Jae-Joon},
  journal = {arXiv preprint arXiv:2610.21207},
  year    = {2026}
}
```