RF-DETR Seg — Segmentation AI Model
RF-DETR Seg Medium returns a separate mask per instance across the 80 COCO categories, so two overlapping people come back as two objects rather than one blob. Real-time, with a DINOv2 backbone and a mask head.
What it does
RF-DETR-Seg-Medium returns four things for every object it finds in a photograph: a label, a box, a confidence score, and a pixel-precise mask of the object's actual shape.
Where a plain detector says there is a dog inside this rectangle, this one gives the outline of the dog. Label, box and mask all come out of a single pass, so no separate segmentation step is needed.
When to use it
- Exact outlines rather than boxes, for background removal and compositing
- Separating two overlapping objects of the same kind into two masks
- Label, box and mask from one run instead of a detector plus a segmenter
- The 80 everyday COCO categories, without any training
- A base to fine-tune on a custom dataset, as Roboflow intended
Details
- CategorySegmentation
- Year2026
- LicenseApache-2.0
Compliance & Provenance
- ProviderRoboflow (open) · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- ImageInputimage
Source image for instance segmentation
- ImageOutputimage
Per-instance masks rendered in distinct colors on a black background
- TextOutputstring
Detected instance labels with their mask color codes and confidence scores
Parameters
- thresholdfloatdefault 0.5range 0–1
Minimum confidence score to keep an instance (higher = fewer, more confident masks).
Tags
Alternatives in Segmentation
- BiRefNet (Subject)
Cleanly extracts a single salient subject with high-resolution edges. ~1-3s.
- SAM2 (Scene)
Segment every object in an image (no labels needed). 10-30s.
- SAM3 (Scene)
Language-prompted segmentation (e.g. "person", "red car"). 10-30s/class.
- SAM3.1 (Scene)
Language-prompted segmentation — SAM 3.1 Object Multiplex with improved accuracy over SAM3. 10-30s/class.