Import, verify, and export annotations

Choose the format that matches the source file instead of converting it implicitly in the dialog.

Select the CSV or JSON file, or select at least one mask for mask import.

Inspect sample, annotation, label, category, image, or color counts shown in the preview.

Complete every required sample-name, label-field, category, or foreground-color mapping before upload.

Choose an optional split only when appropriate. Test is unavailable when the destination splits are locked.

Submit the import, review the result summary, and re-open matched samples before importing another file.

Map the CSV column that contains the name used by the existing dataset samples.

Set each relevant field to Categorical, Continuous, or Segmentation, and leave unrelated fields as None.

Assign imported matches to an available split only when that is part of the intended data policy.

Check mapped fields, match counts, label values, and any skipped rows before submitting.

Open samples from different parts of the CSV and confirm categorical values, numeric values, or segmentation data as applicable.

Choose platform JSON for native sample annotation objects or COCO JSON for a COCO dataset structure.

For platform JSON, review sample, annotation, and label counts. For COCO, review images, annotations, categories, and filename matches.

Correct missing arrays, unknown categories, duplicate identifiers, or filenames that do not match destination samples.

After completion, open representative classifications, boxes, polygons, or other supported geometry and confirm placement and labels.

Select one or more mask files and wait for foreground-color detection to finish.

Use a single label when all foreground belongs to one class, or use color mapping for multi-class masks.

Assign each detected non-background color to the intended dataset label before upload.

Confirm filenames, detected colors, and mapped labels, then submit and inspect mask alignment on representative samples.

Record matched, imported, skipped, unmatched, and failed counts instead of relying only on the success banner.

Use the referenced row, filename, category, field, color, or geometry to repair the input outside the platform.

Open several successfully matched samples and compare them with examples that were skipped or unmatched.

Avoid duplicating successful annotations by limiting the retry to unresolved content whenever the source format allows it.

Choose Download Annotations, select All or one split, select CSV or JSON, and wait for export preparation and browser download.

Use the history to identify previously uploaded annotation sources and their available metadata.

Use the file download action when you need to audit or correct the imported source.

Choose delete only when permitted, read the confirmation for the exact file, and verify the dataset afterward. File-history deletion is a separate operation from editing sample annotations.