TextureBench-3DGS

Note. This is an anonymised copy of the benchmark released for the double-blind review process. It contains no author information; the full release with attribution, licence details and accompanying code will be published after the review.

TextureBench-3DGS is a benchmark of 28 real outdoor scenes with dense natural texture (gravel, grass, leaves, brick, concrete, tiles, foliage), each captured as a multi-view photo set suitable for 3D Gaussian Splatting (3DGS), together with a Poison-Splat poisoned twin of every scene. Clean scenes are high-frequency by nature, so they stress detectors that rely on a high-frequency or anisotropy prior to recognise Poison-Splat inputs. The benchmark is meant for evaluating detectors and purifiers; it was never used to train or calibrate any detector reported alongside it.

Layout

TextureBench/
`-- <scene_id>/
    |-- clean/
    |   |-- images/         # Clean views
    |   |-- sparse/0/       # COLMAP reconstruction
    |   `-- config.yaml     # Capture and preprocessing
    `-- ps_eps16/
        |-- images/         # Poisoned views
        |-- sparse/0/       # COLMAP reconstruction
        `-- config.yaml     # Source images, attack parameters, generation settings

28 scenes, the same file names in clean/images and ps_eps16/images.

clean/

• images/: the released clean views. EXIF removed. These are the images detectors are scored on.
• sparse/0/: COLMAP model (cameras.bin, images.bin, points3D.bin).

To train a 3DGS model on the clean scene, follow the standard 3DGS convert.py route, i.e. undistort first:

colmap image_undistorter --image_path <scene>/clean/images --input_path <scene>/clean/sparse/0 \
    --output_path <scene>/clean_undistorted --output_type COLMAP
python train.py -s <scene>/clean_undistorted -r -1

ps_eps16/

• images/: the Poison-Splat poisoned views, exactly as emitted by the attacker (bounded attack, ε = 16/255, 12 000 attack iterations, proxy model = the clean 3DGS reconstruction). These are the images detectors are scored on for the poisoned class.
• sparse/0/: COLMAP model (cameras.bin, images.bin, points3D.bin).

Loading

from huggingface_hub import snapshot_download
root = snapshot_download(repo_id="anonymous-user-submission/TextureBench-3DGS", repo_type="dataset")

Privacy and licence

All images are original captures of public outdoor spaces; no people are the subject of any view and EXIF metadata (device, location) has been stripped. Released under CC BY-NC 4.0 for research use.