MediPPD annotation protocol
MediPPD annotations locate the bottle-cap reference, induration region, and visible skin findings in PPD skin-test photographs. An image may contain multiple instances and multiple classes. An absent finding has no row for that class.
| ID | Annotation target |
|---|---|
| 0 | Bottle cap used as the image scale reference |
| 1 | Induration region (硬结区域) |
| 2 | Visible blister |
| 3 | Visible area of necrosis |
| 4 | Visible double-ring reaction |
Released labels
Each image in ppd2216_seg/images/<split>/ is paired with a text file of the same stem in ppd2216_seg/labels/<split>/. The split is train or val. Each nonempty row describes one annotated instance. The class IDs match data.yaml.
- Polygon:
class_id x1 y1 x2 y2 ... xn yn, with at least three vertices tracing the annotated region. - Bounding box:
class_id x_center y_center width height, used for instances with a box annotation.
All coordinates are normalized to the image dimensions. Horizontal coordinates and widths are scaled by image width. Vertical coordinates and heights are scaled by image height. The release contains both row types. Preserve the distinction when converting labels to masks or boxes. YOLO segmentation format