README.md

---
license: cc-by-nc-4.0
task_categories:
  - image-to-3d
tags:
  - 3dgs
  - gaussian-splatting
  - adversarial-attack
  - poison-splat
  - remedygs
  - dl3dv
  - 3d-reconstruction
  - security
size_categories:
  - 100K<n<1M
---

# RemedyGSData: Poisoned 3DGS Training Scenes (eps16–eps150)

**RemedyGSData** is the large-scale poisoned image dataset built for
[**RemedyGS: Defend 3D Gaussian Splatting Against Computation Cost Attacks** (CVPR 2026)](https://github.com/Polly-LYP/RemedyGS).

Recent work such as [Poison-splat](https://github.com/jiajun-cheng/Poison-splat) shows that attackers can poison input
images to trigger excessive Gaussian densification, dramatically increasing GPU memory, training time, and rendering
latency — potentially leading to denial-of-service (DoS) in 3DGS-as-a-service systems. RemedyGSData provides
**702 real-world scenes** poisoned by the bounded Poison-splat attacker at **7 perturbation budgets**
(eps ∈ {16, 26, 35, 40, 50, 100, 150}, in units of 1/255), enabling reproducible benchmarking of defenses against
computation-cost attacks on 3D Gaussian Splatting.

All scenes are derived from [DL3DV-10K](https://github.com/DL3DV-10K/Dataset) and are provided in standard
**COLMAP format** (`images/` + `sparse/0/`), drop-in compatible with vanilla 3DGS training.

**Related resources**

| Resource | Link |
|---|---|
| Code (detector + purifier + victim benchmark) | [github.com/Polly-LYP/RemedyGS](https://github.com/Polly-LYP/RemedyGS) |
| Detector checkpoint (`D_ckpt.pth`, poisoned-vs-clean image classifier) | [huggingface.co/ZincL/RemedyGS-detector](https://huggingface.co/ZincL/RemedyGS-detector) |
| Purifier checkpoint (`G_ckpt.pth.tar`, image recovery generator) | included in the GitHub repo (`ckpt/`) |

## Dataset structure

```
eps16~eps150/
├── eps16_<scene_hash>.tar      # 780→600 scenes attacked with eps=16/255
├── eps26_<scene_hash>.tar      # 18 scenes attacked with eps=26/255
├── eps35_<scene_hash>.tar      # 18 scenes, eps=35/255
├── eps40_<scene_hash>.tar      # 18 scenes, eps=40/255
├── eps50_<scene_hash>.tar      # 16 scenes, eps=50/255
├── eps100_<scene_hash>.tar     # 15 scenes, eps=100/255
└── eps150_<scene_hash>.tar     # 17 scenes, eps=150/255
```

Each tar contains one scene folder:

```
epsXX_<scene_hash>/
├── images/            # final poisoned images (frame_00001.png ... frame_NNNNN.png, RGB PNG)
├── images_15000/      # poisoned images at attack iteration 15000
├── images_20000/      # poisoned images at attack iteration 20000
├── images_25000/      # poisoned images at attack iteration 25000
└── sparse/0/          # COLMAP reconstruction: cameras.bin, images.bin, points3D.bin, points3D.ply
```

Attack configuration (Poison-splat bounded attacker): `--adv_iters 30000 --adv_image_search_iters 25`,
using the clean 3DGS model trained for 30k iterations as the proxy. The eps16 split (600 scenes) is the main
benchmark used in the RemedyGS paper; eps26–eps150 form the perturbation-budget ablation (same 20-scene subset,
with 1 scene unavailable for eps50/eps100).

## Statistics

| eps (×1/255) | scene tars | fully complete* | `images/` files |
|---|---|---|---|
| 16  | 600 | 600 | 201,270 |
| 26  | 18  | 14  | 6,047 |
| 35  | 18  | 12  | 6,048 |
| 40  | 18  | 12  | 6,206 |
| 50  | 16  | 10  | 5,288 |
| 100 | 15  | 7   | 5,125 |
| 150 | 17  | 9   | 5,879 |
| **total** | **702** | **664** | **235,863** |

\* "fully complete" = all four image snapshots (`images/`, `images_15000/`, `images_20000/`, `images_25000/`)
contain the same number of frames. Partial scenes have one or more snapshots missing (attack run interrupted);
see `scene_metadata.csv` for per-scene details. Total dataset size: ~975 GB.

## Usage

Download (whole repo or a subset):

```bash
pip install -U huggingface_hub

# everything (~975 GB)
hf download ZincL/RemedyGS --repo-type dataset

# or a single scene
hf download ZincL/RemedyGS eps16~eps150/eps16_<scene_hash>.tar --repo-type dataset
```

Extract a scene and train vanilla 3DGS on it:

```bash
tar -xf eps16~eps150/eps16_<scene_hash>.tar     # -> eps16_<scene_hash>/

python train.py -s eps16_<scene_hash> -m output/eps16_<scene_hash>   # standard 3DGS
```

Evaluate a defense (e.g. RemedyGS) on a poisoned scene. RemedyGS runs **detector first, purifier second** —
the detector (`DetectorNet`, a poisoned-vs-clean image classifier; not to be confused with the GAN
discriminator, which is only used during purifier training) flags poisoned images, and only those are
recovered by the purifier:

```bash
# Step 1 — detector: classify each input image as poisoned (1) or clean (0)
python inference/detector_inference.py \
  --image_dataset_path . --dataset_list scene_list.txt \
  --output_path ./output --load_existing_model_path ckpt/D_ckpt.pth

# Step 2 — purifier: recover benign images from the poisoned ones
python inference/inference.py \
  --image_dataset_path . --dataset_list scene_list.txt \
  --output_path ./output --load_existing_model_path ckpt/G_ckpt.pth.tar
```

## Metadata

`scene_metadata.csv` — one row per scene tar:

| column | description |
|---|---|
| `folder_name` | scene folder / tar name, `epsXX_<scene_hash>` |
| `scene_id` | original DL3DV-10K scene hash |
| `eps` | perturbation budget (×1/255) |
| `source` | origin split (`sampled_100` main benchmark, `1K` ablation) |
| `num_images` | number of frames in `images/` |
| `complete` | whether all four snapshots are present and consistent |

## License & terms

Scenes are derived from **DL3DV-10K**; by using this dataset you must also comply with the original
DL3DV-10K license terms. This dataset is released for **non-commercial research use** (CC BY-NC 4.0).
The poisoned images were generated with the open-source Poison-splat attacker.

## Citation

```bibtex
@article{li2025remedygs,
  title={RemedyGS: Defend 3D Gaussian Splatting Against Computation Cost Attacks},
  author={Li, Yanping and Liu, Zhening and Li, Zijian and Lin, Zehong and Zhang, Jun},
  journal={arXiv preprint arXiv:2511.22147},
  year={2025}
}
```