Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
2.21k
2.21k
End of preview. Expand in Data Studio

Petrolito: Oil-Contaminated Beach Sand Segmentation

Petrolito is an RGB image dataset for binary semantic segmentation of visible contamination on beach sand. It was created to study perception for a mobile beach-cleaning robot whose future task is to locate a contaminated patch and align a collection mechanism with it.

The dataset contains 231 RGB images and 231 binary masks collected across 17 acquisition sessions from a close, downward-looking viewpoint. The masks were manually reviewed and identify pixels corresponding to sand affected by the visual contamination mixture.

This dataset represents visual appearance only. It does not identify the type of petroleum, concentration, toxicity, weathering state, or contamination below the surface.

Key features

  • 231 image-mask pairs.
  • 17 independent acquisition sessions, used to separate train, validation, and test without mixing nearby frames from the same sequence.
  • RGB images originally captured at 2208 x 1242 pixels.
  • A stereo RGB camera mounted on a fixed structure, using a close, downward-looking viewpoint similar to the one intended for the robot.
  • Variation in dry and wet sand, direct sunlight and shade, deposit size, camera motion, rocks, and small debris.
  • Binary task: background versus oil_contaminated_sand.
  • Depth was available during acquisition, but the segmentation benchmark uses RGB images only.

Dataset composition and splits

The splits were defined by acquisition session rather than by individual image:

Split Sessions Images Positive Negative
Train 8 150 130 20
Validation 4 46 46 0
Test 5 35 23 12
Total 17 231 199 32

An image is considered positive when its reviewed mask contains at least one contamination pixel. This definition is used only to summarize the dataset; training and evaluation are pixel-based.

Acquisition and visual contamination mixture

Images were acquired in a 1.5 m x 1.5 m wooden frame filled with beach sand. For prototyping, the visual mixture consisted by mass of 60% cooking oil and 40% black ochre. The mixture produces irregular, visible patches on the sand, but it is not a chemical equivalent of crude or weathered petroleum.

RGB frames were extracted from videos at approximately 1.5-second intervals. The camera was moved around the frame to reproduce different viewing positions and collection passes; some sequences include faster motion and motion blur.

Annotation

Each image was paired with a binary mask. SAM 3 was used in CVAT to generate an initial proposal, after which annotators corrected boundaries and added missed regions. The automatic proposals were therefore annotation aids, not final ground truth.

The main annotation rules were:

  • Sand darkened by partial absorption of the mixture is labeled as contamination.
  • Stains on rocks are excluded because the current task targets contaminated sand.
  • No minimum connected-component size filter is applied.
  • Any mask value greater than zero represents the foreground class.

Because SAM 3 was also evaluated later, some residual spatial bias from its initial proposals may remain despite manual review.

Recommended format

The local project uses the following organization:

data/
β”œβ”€β”€ train/
β”‚   β”œβ”€β”€ image_0001.jpg
β”‚   β”œβ”€β”€ image_0001_mask.png
β”‚   └── ...
β”œβ”€β”€ validation/
β”‚   └── ...
└── test/
    └── ...

Each image should have a mask with the same base name and the _mask suffix. Masks are single-channel images: background is 0, and any value greater than 0 is interpreted as contamination.

If the dataset is loaded with datasets, the equivalent call is:

from datasets import load_dataset

dataset = load_dataset("Aniela13/Petrolito")

The exact split structure and column names depend on how the files are published in the Hugging Face repository.

Intended use

This dataset is intended for:

  • training and comparing semantic-segmentation models;
  • studying close-range visual perception for beach-cleaning robots;
  • analyzing robustness to changes in illumination, moisture, deposit size, and nearby objects;
  • developing future interfaces between an RGB mask, registered depth, and robot navigation.

It should not be treated as a validated autonomous system for detecting real spills, activating machinery, estimating toxicity, or making emergency decisions.

Reference results

As a reference, three task-specific models and SAM 3 in a zero-shot configuration were evaluated on the same 35 test images:

Model Dice/F1 IoU Precision Recall
U-Net 0.9718 0.9452 0.9527 0.9918
DeepLabV3+ 0.9638 0.9301 0.9613 0.9663
SegFormer 0.9610 0.9250 0.9578 0.9643
SAM 3, zero-shot 0.6234 0.4529 0.9759 0.4580

These results are benchmark references, not a guarantee of performance outside the acquisition conditions. In particular, SAM 3 produced precise masks where it detected contamination, but omitted a substantial portion of the labeled regions.

Limitations

  • The dataset is small and was collected in a controlled environment.
  • The visual mixture is not chemically representative of real petroleum.
  • The evaluation does not establish robustness across other beaches, cameras, seasons, distances, or weather conditions.
  • Labels describe visible contamination on sand; they do not include concentration, depth, hydrocarbon type, or a complete taxonomy of contaminated rocks and debris.
  • The camera was mounted on a structure during acquisition rather than directly on the mobile robot.
  • A closed-loop connection between perception, navigation, and the collection mechanism has not yet been implemented or validated.
  • The test split contains 35 images, so the metrics should be interpreted with caution.

License and responsible use

Check the license specified in the repository configuration before redistributing or using this dataset. If no explicit license is provided, contact the authors before reusing it in a product or operational system.

Any model trained with this dataset should be validated on independent data and under human supervision before use in real environmental-cleanup or robotic control scenarios.

Citation

If you use this dataset, please cite the Hugging Face repository:

@dataset{petrolito,
  author    = {Aniela13},
  title     = {Petrolito: Oil-Contaminated Beach Sand Segmentation},
  publisher = {Hugging Face},
  year      = {2026},
  url       = {https://huggingface.co/datasets/Aniela13/Petrolito}
}

For the robotic context and complete benchmark, also consult the associated Petrolito paper and code.

Downloads last month
991