Supported tasks and formats
Match the annotation meaning to the task you intend to evaluate.
A format describes how data is stored. A task describes what predictions and reference labels mean. Check both when importing.
Image formats
| Source | Supported annotation use |
|---|---|
| COCO JSON / RLE | Boxes, polygons, instance masks; explicit outlines can convert to oriented boxes |
| YOLO TXT / YAML | Boxes, polygons, oriented boxes |
| Class folders / classification manifests | One class per image |
| Pascal VOC XML | Boxes and supported polygon extensions |
| LabelMe JSON | Rectangles and polygons; grouped parts retain one instance |
| CreateML | Detection manifests and referenced images |
| Open Images CSV | Bounding boxes with class descriptions and retained source flags |
| RetinaNet Keras CSV | Headerless image,x1,y1,x2,y2,class_name rows and retained images |
COCO/YOLO box and mask support does not imply keypoint support. Box-to-mask and polygon-to-box conversions are offered explicitly; review their effect on the labels.
YOLO sources
data.yaml and dataset.yaml can name class lists or indexed maps and split folders or image lists. References must stay inside the inspected source. Empty labels establish known negatives; missing labels retain unfinished annotation status.
Tables
CSV, TSV, and Parquet support tabular classification and regression. Review column mapping and the target before importing.
Import the Dataset, then choose a compatible model.