> For the complete documentation index, see [llms.txt](https://docs.roboflow.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.roboflow.com/workflows/developer-guide/developer-guide/kinds/semantic-segmentation-prediction.md).

# semanticsegmentationprediction Kind

Prediction with per-pixel class label and confidence for semantic segmentation

## Data representation

{% hint style="warning" %}
**Data representation**

This kind has a different internal and external representation. **External** representation is relevant for integration with your workflow, whereas **internal** one is an implementation detail useful for Workflows blocks development.
{% endhint %}

### External

External data representation is relevant for Workflows clients - it dictates what is the input and output format of data.

Type: `dict`

### Internal

Internal data representation is relevant for Workflows blocks creators - this is the type that will be provided by Execution Engine in runtime to the block that consumes input of this kind.

Type: `sv.Detections`

## Details

This kind represents a single semantic segmentation prediction as an [`sv.Detections(...)`](https://supervision.roboflow.com/latest/detection/core/) object with one detection per predicted class. Each detection carries an RLE-encoded mask covering all pixels assigned to that class.

**Why RLE and not polygons:**

Semantic segmentation assigns a class label to every pixel in the image. A single class can appear in multiple spatially disconnected regions (e.g., two separate "person" regions on opposite sides of the frame). Polygon-based serialization uses `cv2.findContours()`, which only retains the first contiguous contour and silently discards all others - causing irreversible data loss for non-contiguous masks. RLE (Run-Length Encoding, COCO standard) is a pixel-level encoding that represents the complete mask regardless of spatial topology, making it the only correct serialization format for semantic segmentation masks.

**Internal representation:** `sv.Detections` with:

* `xyxy` - tight bounding box enclosing all pixels of the class
* `class_id` - integer class ID
* `confidence` - mean confidence over all pixels of the class
* `data["class_name"]` - class label string
* `data["rle_mask"]` - numpy object array of COCO RLE dicts `{"size": [H, W], "counts": "..."}`

**Serialised format** (one entry per class in `predictions`):

```json
{
    "image": {"width": 640, "height": 480},
    "predictions": [
        {
            "x": 320.0, "y": 240.0, "width": 200.0, "height": 180.0,
            "confidence": 0.92,
            "class_id": 1,
            "class": "person",
            "detection_id": "a1b2c3d4-...",
            "rle_mask": {"size": [480, 640], "counts": "XYZ..."}
        }
    ]
}
```

**Decoding RLE masks:**

```python
import pycocotools.mask as mask_utils
import numpy as np

rle = prediction["rle_mask"]
binary_mask = mask_utils.decode(rle).astype(bool)  # shape: (H, W)
```