Glossary¶
Quick reference for the terms and abbreviations used throughout the LightlyStudio documentation.
| Term | Description |
|---|---|
| Annotation | A classification, object-detection box, or segmentation mask attached to a sample. Each annotation has an annotation class and an annotation type, and belongs to an annotation source. Can represent ground truth or a model prediction. LightlyStudio uses "annotation" rather than "label" for this concept. See Annotations. |
| Annotation class | The category assigned to an annotation, such as dog, cat, or car. Not to be confused with annotation type, which describes the form of the annotation. |
| Annotation crop | The image region defined by an annotation's bounding box or mask, extracted as a standalone image. Used to compute per-annotation embeddings. |
| Annotation source | A named group of annotations, such as ground_truth or model_v1. Keeps annotations from different people, tools, or models separate. Compare with tag, which groups samples rather than annotations. |
| Annotation type | The form of an annotation: classification, object detection, or segmentation. Compare with annotation class, which describes the category. |
| Caption | A text description associated with a sample, for example as training data for a vision-language model. See Captions. |
| Dataset | A collection of samples managed together, including their annotations, captions, embeddings, metadata, and tags. A dataset can be an image dataset or a video dataset. |
| Dataset distribution | A summary of how annotations or metadata values occur across the current view. See Dataset Distributions. |
| Embedding plot | An interactive 2D projection of embedding space where each point represents a sample or annotation. See Embeddings. |
| Evaluation run | A saved comparison between a ground-truth and a prediction annotation source. Stores configuration, results, a confusion matrix, and per-sample metrics. See Model Evaluation. |
| Export | A copy of dataset content saved in a standard format (such as COCO or YOLO) for use outside LightlyStudio. See Export. |
| Frame | A still image extracted from a video. LightlyStudio retains its frame number, timestamp, and parent video. Each frame is a sample. See Video Dataset. |
| Image dataset | A dataset whose primary samples are images. See Image Dataset. |
| Lightly Query Language | SQL-like language used by the GUI query editor to filter images by fields, tags, and annotations. See LQL reference. |
| Plugin | An installable extension that adds a workflow to LightlyStudio, such as model-assisted segmentation. See Plugins. |
| Python API | The classes and functions in the lightly_studio package for working with datasets programmatically. See Dataset API. |
| Query | A reusable description of a dataset subset. In Python, a query can filter, sort, and slice samples. In the GUI, it uses Lightly Query Language. See Query in Python. |
| Sample | An individual data item — an image, a video, or an extracted video frame. Samples are the basic unit of a dataset and can carry annotations, captions, embeddings, metadata, and tags. |
| Sampling | Algorithmic selection of a smaller, more useful subset from a dataset. The result is stored as a tag. See Sampling. |
| Similarity search | Embedding-based search that ranks samples by visual or semantic similarity to a given input in embedding space. See Search in the GUI. |
| Tag | A named marker for organizing samples into reusable groups, such as reviewed or train. Sampling results are saved as tags. Compare with annotation source, which groups annotations rather than samples. See Tags. |
| Video dataset | A dataset whose primary samples are videos. Supports frame extraction and frame-level annotations. See Video Dataset. |
| View | A filtered or sorted perspective on a dataset that shows a subset of its samples without modifying the underlying data. Filters and queries produce views. |
