Noktura

Annotations

Upload labels for the images in your datasets, then browse and filter by class across the platform.

Last updated: May 2026

What annotations are

An annotation is a label drawn on an image — usually a bounding box or polygon around a specific plant or object. Noktura imports labels you've already made in another tool (CVAT, Roboflow, Labelbox, etc.) and lets you search across all your annotated images.

Supported formats

  • COCO — single JSON file. Bounding boxes, polygons, RLE crowd masks.
  • WeedCOCO — COCO with extra AgContext fields (BBCH, agent, growth stage).
  • YOLO — zip containing data.yaml and per-image .txt files. Both YOLO-bbox and YOLO-seg supported.
  • Pascal VOC — single XML or a zip of XMLs.

Up to 500 MB per file. Larger label sets can be split across multiple uploads.

Uploading labels

1

Open your dataset

Annotations attach to a specific dataset, so the images need to exist first. Open the dataset you want to label, or create one via Uploading Datasets.

2

Open the annotations workspace

Click the Annotations button in the dataset header, or head to Annotations in the left sidebar to see all your datasets in one place.

3

Upload your file

Click Upload annotations and drop your file. Noktura detects the format automatically from the file's contents.

4

Map classes

If your file has classes Noktura can read ahead of time (COCO, YOLO), you'll see a quick mapping step. Class names that exactly match existing ones in your dataset are merged automatically; new names become new classes. You can also skip a class to leave its annotations out.

5

Wait for parsing

Once uploaded, parsing happens in the background. Small files finish in seconds; large files can take a few minutes. You can leave the page and come back — the status shows on the dataset's annotations page.

Tip

Labels are matched to your images by filename. If your YOLO file's labels/dog_001.txt needs to land on dog_001.jpg in your dataset, make sure the filenames match (extension aside).

Re-uploading a corrected file

If you find errors in your labels and re-export from your annotation tool, upload the corrected file. Noktura asks whether to replace your earlier upload or add the new file alongside it. Replacing is the default — the older version is removed cleanly.

Annotations from other contributors on the same dataset are never affected by your replace action.

Browsing annotated images

Use Browse annotated images on the annotations page (or go to Explore → Images) to search across every dataset you can see. Filter by:

  • Annotated species (taxonomy-aware via GBIF)
  • Annotation density (how many labels per image)
  • License, crop type, date range, region, weather

Who can upload

  • Dataset owners can always upload.
  • Group admins can upload to datasets shared with their group.
  • Regular group members can view annotations but not modify them.

Downloading annotated images

Downloading a curated subset of annotated images across datasets is on the roadmap. Today you can download individual datasets and their annotations from the dataset detail page.

Troubleshooting

Annotations show up under “referenced images not in this dataset”

This means filenames in your annotation file don't match images in the dataset. Common causes: a directory prefix in the filename, an extension mismatch (.jpg vs .JPG), or labels from a different image set entirely.

Format not detected

Make sure your file is one of the four supported formats. For zips, the inside layout matters: YOLO needs a data.yaml + a labels/ directory; Pascal VOC needs .xml files in a recognisable folder structure.

Upload failed mid-parse

Use Retry on the failed upload — the file is kept for 30 days. If retries keep failing, the error message on the upload row will point at the underlying cause (most often a malformed annotation or invalid bounding-box coordinates).

Next steps