Model
YOLO11-seg
Produce per-pixel masks that separate standing objects from the background plate.
beta
Segmentation
MasksForeground
Overview
A YOLO11 segmentation head yields instance / foreground masks. Overseer combines them with a depth-derived foreground mask to decide which pixels are 'standing objects', so scene completion can inpaint depth and texture behind them and reconstruct the occluded background as a real surface.
| Role | Segmentation |
|---|---|
| Expected latency | ~12–20 ms / frame (GPU) |
| Hardware requirements | NVIDIA GPU recommended |
| Memory requirements | ~2 GB VRAM |
Inputs & Outputs
Inputs
- RGB frame
Outputs
- Instance masks
- Foreground / standing-object mask
Advantages & Limitations
Advantages
- Clean mover / background separation
- Improves reconstruction seams
- Shares the YOLO backbone
Limitations
- Mask edges soften on motion blur
- Thin structures under-segmented
- Beta: not yet wired into every analytic
Example outputs
YOLO11-seg
A person mask lifted off the wall behind them
YOLO11-seg
A foreground mask driving background inpainting
Configuration
Keys in config/default.yaml that govern this model.
| Key | Default |
|---|---|
spatial.complete | true |
detectors.seg.model | yolo11n-seg.pt |
Benchmarks
Placeholder
Benchmark figures below are illustrative and awaiting a standardised harness. Treat them as placeholders.
| Benchmark | Dataset | Metric | Value | Hardware |
|---|---|---|---|---|
| Mask mAP | COCO (ref) | mAP | ~40 | reference |