Labelling Quality Test

Check annotation accuracy in datasets. Improve object detection results with cleaner labels.

The Labelling Quality Test highlights data points with a higher probability of labelling errors. By setting a threshold on the provided mistake score metric, you can identify and rectify labelling inaccuracies.

Execute Test:

The following code snippet is aimed at performing a Labelling Quality Test on a specified dataset to evaluate the quality of the labelling done.

rules = LQRules()
rules.add(metric="mistake_score", label=["ALL"], metric_threshold=0.5)


edge_case_detection = labelling_quality_test(test_session=test_session,
                                             dataset_name="labeling_all_the_instance_mar_18",
                                             test_name="Labeling Quality Test",
                                             type="labelling_quality",
                                             output_type="instance_segmentation",
                                             mistake_score_col_name="MistakeScores",
                                             gt="GT",
                                             rules=rules)

test_session.add(edge_case_detection)
test_session.run()

• LQRules(): Initialises the labelling quality rules.

• rules.add(): Adds a new rule for assessing labelling quality:

  • metric: The performance metric to evaluate, "mistake_score" in this case.

  • label: Specifies the label(s) these metrics apply to. ["ALL"] means all labels in the dataset.

  • metric_threshold: The threshold for the mistake score, above which the label is considered incorrect.

• labelling_quality_test(): Prepares the labelling quality test with the following parameters:

  • test_session: The current session linked to your project.

  • dataset_name: The name of the dataset to be evaluated.

  • type: The type of test, "labelling_quality" here, which focuses on how consistently the labelling is done.

  • output_type: The type of output expected, "instance_segmentation" in this context.

  • mistake_score_col_name: The column in the dataset that contains the mistake scores.

  • rules: The previously defined rules for the test.

test_session.add(): Registers the labelling quality test with the session.

test_session.run(): Starts the execution of all tests in the session, including your labelling quality test.

By following the steps outlined above, you have successfully set up a Labelling Quality Test in RagaAI.

Analysing Test Results

Understanding Mistake Score

• Mistake Score Metric: A quantitative measure indicating the likelihood of labelling errors in your dataset.

Test Overview

• Pie Chart Overview: Shows the proportion of labels that passed or failed based on the Mistake Score threshold.

Mistake Score Distribution

• Bar Graph Visualisation: Displays average Mistake Scores for failed labels, class-wise, and the volume of failed data points per class.

Interpreting Results

• Passed Data Points: Identified by meeting the Mistake Score threshold, indicating accurate labelling.

• Failed Data Points: Exceeding the threshold, suggesting potential labelling inaccuracies.

Visualisation and Assessment

• Visualising Annotations: Arranges images by descending Mistake Score for label assessment.

Image View

• In-Depth Analysis: Analyse Mistake Scores for each label in an image, with interactive features for annotations and original image viewing.

• Information Card: Provides details like Mistake Score, threshold, area percentage, and confidence score for each label.

Note: Mistake Scores are not calculated for annotations covering less than 5% of an image.

By following these steps, you can effectively utilise the Labelling Quality Test to identify and address labelling inaccuracies in your datasets, enhancing the overall quality and reliability of your models.