---
task_categories:
- image-segmentation
tags:
- semantic-segmentation
- capybara
- computer-vision
- synthetic-data
pretty_name: OpenCAPY
size_categories:
- 10K<n<100K
---

# OpenCAPY

OpenCAPY is a capybara semantic segmentation dataset.

## Dataset Summary

OpenCAPY provides RGB images and single-channel PNG segmentation masks for capybara segmentation. Mask pixel values are class indices, not visualization values.

- Task: semantic segmentation
- Images: RGB image files under `images/`
- Masks: single-channel PNG files under `labels/`
- Label value `0`: background
- Label value `1`: capybara
- Training data includes real and synthetic data
- Validation and test data are real data

## Dataset Structure

```text
OpenCAPY/
├── README.md
├── RELEASE_MANIFEST.txt
├── RELEASE_REPORT.md
├── dataset.json
├── images/
├── labels/
├── metadata/
└── splits/
```

Each image, label, and metadata file uses the same relative stem. For example:

```text
images/train_real/night/img_xxx.jpg
labels/train_real/night/img_xxx.png
metadata/train_real/night/img_xxx.json
```

Split files contain extensionless image ids, for example:

```text
train_real/night/img_xxx
```

The `dataset.json` file lists samples with relative paths for images, labels, and metadata.

## Dataset Splits

```text
train_synthetic_qc_cleaned   23626
train_real                    6494
val_real                       390
test_real                      382
----------------------------------
total                        30892
```

## Labels

Masks are single-channel PNG files with class-index values:

```text
0 background
1 capybara
```

Do not scale training labels to `0/255`. Visualization masks and preview overlays should be generated outside the canonical dataset files.

## Data Sources

The release contains both real and synthetic capybara image subsets. The dataset metadata and split names distinguish real subsets from the `train_synthetic_qc_cleaned` synthetic training subset.

Confirmed subset names in this release:

- `train_real`
- `train_synthetic_qc_cleaned`
- `val_real`
- `test_real`

## Annotation / Processing

The dataset contains segmentation masks stored as canonical `0/1` class-index PNG files. Repository metadata indicates that automated processing used Qwen grounding and SAM3 segmentation for real subsets, and that synthetic training data passed a QC-cleaned pipeline before inclusion.

This release packaging did not rerun annotation, regenerate masks, change split membership, or perform a new data-quality audit. Sample identity and split identity were preserved from the frozen source dataset.

## Intended Uses

- Capybara semantic segmentation
- Computer vision research
- Segmentation benchmarking
- Real/synthetic training experiments
- Robustness and domain-gap studies involving animal segmentation

## Limitations

- The dataset is focused on capybaras and may not generalize to other animals or object categories.
- Synthetic images may have a domain gap relative to real-world imagery.
- Data sources and visual distributions may contain biases.
- Users should validate generalization for their own deployment or research setting.

## License / Redistribution

License and redistribution terms are being finalized before public release.

The final dataset license and original image redistribution status must be confirmed by the publisher before making the Hugging Face repository public.