Annotate

Claude vision annotation correction using multi-pass refinement.

STATUS: EXPERIMENTAL

Approach

1. YOLO inference at low confidence (0.15) → precise boxes
2. Claude vision review → keeps good, removes bad, identifies missed
3. For missed objects → YOLO on cropped region
4. Final Claude validation

Workflow

1. Read yolo-project.yaml for class names
2. For each image: a. Draw existing annotations: python scripts/draw_annotations.py b. Claude reads annotated image c. Evaluate: keep / correct / remove / add d. Write corrected YOLO label