sailswarm-groundingdino-labeler

SailSwarm GroundingDINO labeller (config / recipe — no weights)

This repo ships a configuration, not new weights. It is the tuned open-vocabulary detection recipe used as the labelling oracle for the SailSwarm obstacle-detection module — it pre-fills bounding-box suggestions that are human-audited and then distilled into the deployable YOLOv8n.

The model itself is the unmodified IDEA-Research/grounding-dino-base; we only tuned the prompts, thresholds, and a multi-pass merge.

Recipe (detector_labeler.yaml)

Multi-pass, because GroundingDINO degrades with long prompts — each pass has a short focused phrase list and its own thresholds, merged with per-class NMS:

• objects (box=0.30): boat / duck / buoy / person on the water.
• structure (box=0.20): poles / piling / ladder / platform.
• structures_large (box=0.20): crane / gantry / machinery.
• image_shortest_edge: 1000 (recovers thin poles + small distant objects); boat_gate: 0.8 (drops mast/rig boxes inside a boat to lift structure precision).

Result

Class-agnostic F1 0.701 (P 0.76, R 0.65) on the held-out Konstanz eval clip — the strongest box source we have; the distilled YOLOv8n reaches F1 0.673 and is the Pi-deployable detector.

Use

python -m scripts.eval.detector_labeler (SailSwarm repo) with this config.

License

GroundingDINO base model: Apache-2.0 (IDEA-Research). This config: same.