Lightly

O que é Lightly?

Lightly is a computer vision platform for ML engineers and organizations managing visual datasets and developing models. It connects data selection, annotation, quality review and model evaluation in a shared workflow.

Its two core products serve different stages of development: LightlyStudio provides dataset curation, labeling, visualization and management, while LightlyTrain supports self-supervised pretraining. Teams can explore their data, select samples for annotation and inspect model errors before preparing datasets for downstream training.

O que Lightly pode fazer?

Embedding search and sampling

Indexed data receives automatic embeddings. Combine metadata filters with text or image search, then sample for diversity, deduplication, outlier detection, similarity or class balance.

Annotation and quality review

Create, edit and review annotations in the same workspace. Labeled objects receive individual embeddings for object-level search and clustering; annotations are versioned.

Model error analysis

Evaluate detection, classification and segmentation against ground truth. Inspect per-sample metrics and a confusion matrix, then sort by false positives, false negatives or mean intersection over union.

Python plugins

Install plugins for SAM3 segmentation, LightlyTrain auto-labeling and video bounding box propagation. Register custom Python functions with forms so colleagues can run them from the interface.

Team access controls

Assign Viewer, Labeler, Editor or Admin roles to control browsing, annotation, workspace configuration and administration. Cloud credentials can be managed centrally.

Dataset export

Export annotations or filenames through the interface or Python. Choose a full dataset or a subset filtered by tags or queries, with formats including COCO, YOLO and Pascal VOC.

Casos de uso

“Prioritize a labeling backlog”

For a large collection of raw images, select a varied subset and remove duplicates before assigning annotation work, focusing labeling effort on samples that expand dataset coverage.

“Investigate model failure patterns”

After evaluating a model, inspect its worst predictions to identify recurring errors and decide which examples need annotation review or additional training coverage.

“Explore egocentric video datasets”

Organize action clips with captions and metadata for robotics research. Lightly demonstrates this workflow with more than 37,000 EPIC-KITCHENS-100 clips, using embedding plots, text search and diversity sampling.

Perguntas frequentes

How do LightlyStudio and LightlyTrain differ?

LightlyStudio handles dataset exploration, curation, annotation and evaluation. LightlyTrain focuses on self-supervised pretraining for computer vision models.

Can teams automate labeling?

LightlyStudio supports SAM3 and LightlyTrain auto-labeling plugins, alongside manual annotation and review. A separate plugin supports propagating video bounding boxes.

Can Lightly fit an existing ML workflow?

Lightly provides Python interfaces for custom plugins and dataset exports. Its supported export formats let teams prepare annotations for downstream training tools.

Is there a free starting option?

The website invites users to try LightlyStudio for free and offers demo bookings.