Segment Image (RF DETR)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is a single downstream-resolved choice fixed at deploy time; every arm reprojects the same instances, so masks, boxes and classes stay aligned.
• Hard constraintAn RF-DETR-Seg weights file is required, and the backbone-variant setting must match the backbone the checkpoint was trained with or loading fails on a state-dict mismatch; a detection-only checkpoint produces no instance masks.
• Parameter interactionThe class-id allow-list is applied after inference, so the model still runs the full forward over every input; the per-class confidence overrides set thresholds for specific classes with the rest falling back to the default confidence threshold.
• CompatibilityClass ids depend on the loaded checkpoint, so downstream rules that branch on a class id must be re-checked when the weights change.
• CostA GPU is strongly preferred; transformer segmentation on CPU is slow for live streams.
• State lifetimeThe model, the backbone variant and all configuration are captured once at startup; replacing the checkpoint requires a redeploy.