---
license: other
license_name: waymo-dataset-license-non-commercial
license_link: https://waymo.com/open/terms/
library_name: pytorch
tags:
  - gaussian-splatting
  - novel-view-synthesis
  - autonomous-driving
  - waymo-open-dataset
  - armgs
datasets:
  - waymo-open-dataset
---

# Ours_Waymo_ArmGS

ArmGS checkpoints trained for dynamic urban novel-view synthesis on the Waymo Open Dataset.

## License and required notice

This model was made using the Waymo Open Dataset, provided by Waymo LLC under the Waymo Dataset License Agreement for Non-Commercial Use. Access, use, redistribution, and modification of this model are governed by that agreement, including its non-commercial restrictions.

Read [WAYMO_DATASET_LICENSE_NOTICE.md](WAYMO_DATASET_LICENSE_NOTICE.md), the included archived agreement, and the [current official terms](https://waymo.com/open/terms/) before downloading or using these files. These checkpoints must not be used in vehicle operation, production systems, or primarily commercial applications.

## Release contents

- Nine completed 30,000-step training-split runs under `waymo/<sequence>/splatad_30k/`.
- One completed 30,000-step validation reference under `waymo/10448102132863604198_472_000_492_000/paper_reference_30k/`.
- Every released run includes `checkpoints/final.pt`, the resolved YAML, run metadata, W&B run identity, and final novel-view/reconstruction metric JSON files.
- `7566697458525030390_1440_000_1460_000` is not included yet because its 30,000-step training was still running when this release was packaged.

The nine training-split runs use the `streetgs-periodic` split, `PAPER_MODE=0`, and Waymo GT `lidar_box` actor tracking fallback. They must not be described as official SplatAD LINSPACE50 results. The separate validation reference uses CAStrack and centered known-pose COLMAP preprocessing.

See `release_manifest.json` for exact sequence IDs, checkpoint sizes, W&B IDs, and protocol metadata.

## Loading

These are full PyTorch trainer checkpoints produced by the ArmGS implementation, not standalone `safetensors` weights. Use the matching ArmGS code and the included resolved config. As with any pickle-based PyTorch checkpoint, only load files obtained from a trusted source.

```python
import torch

checkpoint = torch.load("waymo/<sequence>/splatad_30k/checkpoints/final.pt", map_location="cpu")
print(checkpoint["trainer"]["step"])
```

## Citation

```bibtex
@misc{waymo_open_dataset,
  title   = {Waymo Open Dataset: An autonomous driving dataset},
  website = {https://www.waymo.com/open},
  year    = {2019--2025}
}
```