Fix classification labels

Scenario: we have a classification dataset with a few labels that are not correct. We want to fix them. This classification dataset has been formed from another detection dataset which contains original labels. That's why we want to find respective detection labels in the original dataset, fix them and re-create our classification dataset.

We will use the following datasets: - classification_dataset as a classification dataset, which contains a sample tag with the name of the original dataset. It also contains a label field class_label with classification labels. - detection_dataset as a detection dataset. It contains labels in a detection field detection.

Steps: 1. Mark wrong labels in the classification dataset from the FiftyOne App. 2. Print the wrong labels in the classification dataset. 3. Assign a label tag to corresponding detection labels in the original dataset. 4. Send to the annotation server. 5. Annotate incorrect labels. 6. Fetch updated labels 7. Re-create the classification dataset without wrong labels.

Solution

Mark wrong labels.

Run fiftyone app launch and open classification_dataset in the FiftyOne App. Select wrong labels and mark them with fix_labels sample tag.

Print wrong labels

finegrained data display print_labels \
  --dataset classification_dataset \
  --label-field class_label \
  --include-tags fix_labels > fix_labels.txt

This will print all wrong labels in the fix_labels.txt file. Open the file and see its contents. It should contain wrong labels selected in the previous step.

Assign a label tag to detection labels

finegrained data tag tag_labels \
  --dataset detection_dataset \
  --label-field detection \
  --label-tags fix_labels \
  --labels fix_labels.txt

This will read labels from the file created at a step before filter the detection dataset to only contain those labels and assign a label tag to each detection label. Review results in the FiftyOne App by navigating to detection_dataset and filtering with label tag fix_labels.

Send to annotation server

finegrained data annotations annotate \
 --dataset detection_dataset \
 --annotation-key fix_errors \
 --label-field detection \
  --dataset-kwargs='{label_tags:"fix_labels"}' \
  --backend cvat_server.txt \
  --segment-size 10 \
  --project-id 1

This assumes that you have a CVAT server running and its connection parameters are defined in a cvat_server.txt file as following:

backend=cvat
url=http://<host>:<port>
username=<username>
password=<password>

It also expects a project with id 1 to exist in the CVAT server.

All samples containing labels with the fix-dupl tag will be sent to the CVAT server.

Fix incorrect annotations

Navigate to your CVAT server and open an annotation task. Use a filter on top-right corner to filter by your labels from fix_labels.txt to avoid updating other correct labels. Go through each task and fix incorrect labels.

Once finished, hit "Save" button and mark the task as finished.

Fetch updated labels

finegrained annotations load 
    --dataset detection_dataset  
    --annotation-key fix_errors 
    --backend ./cvat_server.txt 
    --dest-field detection

This will fetch updated labels from the CVAT server and save them to the detection field of the detection_dataset dataset.

If you open the dataset in the FiftyOne App, you will see that the labels have been updated.