README.md

# NSFW Segmentation

Multi-head release of single-task segmentation models targeting NSFW anatomy. Each checkpoint runs independently and produces binary masks for the specified classes.

| File | Backbone | Task | Classes | Mask mAP@0.5 | Mask mAP@0.5:0.95 |
| --- | --- | --- | --- | --- | --- |
| `nsfw-seg-breast-s.pt` | YOLO11s | Breast anatomy | breast, areola, nipple | 0.895 | 0.636 |
| `nsfw-seg-breast-x.pt` | YOLO11x | Breast anatomy | breast, areola, nipple | 0.929 | 0.702 |
| `nsfw-seg-vagina-s.pt` | YOLO11s | Vagina | vagina | 0.995 | 0.871 |
| `nsfw-seg-vagina-x.pt` | YOLO11x | Vagina | vagina | 0.995 | 0.918 |
| `nsfw-seg-penis-s.pt`  | YOLO11s | Penis | penis | 0.995 | 0.975 |
| `nsfw-seg-penis-x.pt`  | YOLO11x | Penis | penis | 0.995 | 0.987 |

## Description
- Backbones: YOLO11s and YOLO11x segmentation heads (Ultralytics 8.3.204).
- Weights exported as `.pt` checkpoints compatible with `ultralytics>=8.3`.
- One model per label space; load the checkpoint that matches your target anatomy.

## Intended Use
- Automatic instance segmentation for NSFW anatomical structures in moderated, research, or medical-support workflows.
- **Inputs:** RGB images.
- **Outputs:** Binary masks aligned with the class taxonomy above.

## Data Summary
- Training data consisted of curated, privately-held NSFW image sets with polygon masks (YOLO segmentation format).
- Train/validation splits were normalized and merged after preprocessing; metrics reflect held-out validation imagery.
- Datasets are not included in this release.

## Metrics
- Evaluated with `yolo segment val` at 832 px resolution, confidence threshold 0.1.
- Numbers in the table refer to the best checkpoint per task.

## Limitations
- Models are not a substitute for clinical assessment.
- Domain shift (lighting, camera quality, demographics) may impact performance.
- No safety filtering is applied; downstream systems must implement access controls.

## Quickstart
```python
from ultralytics import YOLO

model = YOLO("nsfw-seg-breast-s.pt")  # swap for -x or other anatomy
results = model.predict("path/to/image.jpg", imgsz=832, conf=0.1)
```

## Support
For integration questions or feedback, open an issue on the hosting repository and mention the checkpoint name in the subject line.