Annotation file reference and troubleshooting

Choose an annotation import format#

• CSV: map the sample-name field and each label column as categorical, continuous, or segmentation; optionally map split membership.
• Platform JSON: use an array of sample annotation objects matched by name, with an `annotations` array for each sample.
• Spatial platform JSON: boxes and polygons can include frame, slice, and view context. Polygon points use the supported flat coordinate representation.
• COCO JSON: include `images`, `annotations`, and `categories` arrays; image filenames must match uploaded sample filenames.
• Masks: map detected colors to dataset labels, or select a single mask label for the import.
• Imports merge into matched samples. Test consequential mappings on a small clone or representative subset first.

Validate files before import#

• Use stable, unique sample names and match filename spelling and case exactly.
• Keep label names, capitalization, and value types consistent across the dataset.
• Validate coordinates, image dimensions, polygon point order, and class/category identifiers.
• Represent empty annotations intentionally rather than with malformed placeholder geometry.
• For video or volumes, verify frame, slice, and view indices use the expected convention.
• Retain a source export before a large merge or mapping change.

Recover from an import failure#

1. Read the preview and error summary

   Separate unmatched images, invalid annotations, skipped rows, and mapping errors.

2. Fix the source file

   Correct the referenced filenames, columns, categories, dimensions, or geometry outside the import dialog.

3. Retry a small batch

   Confirm representative classification values and spatial shapes before importing the full file.

4. Open matched samples

   Verify labels, boxes, polygons, masks, frame/slice/view, and split membership.

Troubleshoot common workflow problems#

1. An action is missing

   Confirm the active workspace, ownership/collaborator role, organization role, plan capacity, verified email, data type, resource state, and enabled service.

2. Upload processing does not finish

   Keep the queue visible long enough to distinguish browser transfer from server-side processing, then refresh the dataset count once.

3. Annotation assistance is unavailable

   Check the annotation-engine status, supported data type, selected geometry, and Health page before retrying.

4. A pipeline cannot start

   Review dataset readiness, required splits, generated resources, email verification, active duplicate run, and compute capacity.

5. Training runs out of memory

   Use the recommended smaller batch size when provided, or select a compatible architecture/input configuration.

6. Results or previews do not load

   Confirm the selected model version completed, retry the protected artifact once, and review service health.

Preserve data quality and request support safely#

• Visually inspect representative samples from every label and split.
• Prevent subject or patient leakage across training, validation, and test data.
• Record the dataset and model version used for any reported result without copying private identifiers into public material.
• When requesting support, include only the resource type, safe visible status, approximate time, and exact non-sensitive error text.