Skip to content

Visualize Dataset Distributions in LightlyStudio

Dataset distributions summarize the samples and annotations in your current dataset view. Use them to spot class imbalance, coverage gaps, and missing metadata before training or evaluating a model.

Explore Distributions in the GUI

Open an image dataset and click Distr to open the Distribution panel. Choose Annotation classes or Metadata, then select the annotation type or metadata key you want to inspect.

The plots respond to active filters, so you can inspect the complete dataset or a filtered subset.

Annotation Class Distributions

The annotation class distribution shows how annotations are distributed across classes. It supports classification, object detection, and segmentation annotations.

Use Annotation type to inspect one annotation type or all available types together. In the chart settings, use Count by to choose what the bars represent:

  • Objects counts individual annotations, such as bounding boxes or segmentation masks.
  • Samples counts distinct images containing each annotation class.

You can also choose which classes to show, sort them by count or name, change the chart orientation, and expand the chart. These views help identify dominant, rare, and long-tail classes.

Metadata Distributions

Select Metadata, then choose the metadata key you want to inspect.

Numeric Metadata

Numeric metadata fields, such as temperature, brightness, confidence, or blur score, are shown as histograms.

Change the number of bins to adjust the level of detail. Click a bin or drag across multiple bins to filter the dataset to that range. Select the same range again to clear the filter.

Categorical Metadata

Categorical metadata fields, such as city, camera ID, production line, weather, or a boolean review status, are shown as bar charts.

Select one or more values to filter the dataset. Missing represents samples without a value for the selected key. Other groups values outside the most frequent values shown in the chart and cannot be selected.

Required Dataset Content

The available distributions depend on what your dataset contains:

Distribution Required content
Annotation classes Classification, object detection, or segmentation annotations
Numeric metadata Integer or floating-point sample metadata
Categorical metadata String or boolean sample metadata

If a distribution is not available, add the corresponding annotations or metadata to the dataset. See Search and Filter for the other ways to narrow down the samples shown in the dataset view.