The Catalog is a data curation tool for organizing, searching, visualizing, and exploring labeled and unlabeled data (including any metadata). Teams developing and operating production AI systems need a data catalog to enable data selection for downstream data-centric workflows. This includes data labeling, model training, model evaluation, error analysis, and active learning.

Anatomy of the Catalog

Understanding data flow

Data can flow into and out of the Catalog in numerous ways. This diagram indicates how data can flow through the Catalog.

Data inflow

Catalog can ingest the following pieces of data. Catalog makes it easy to search, visualize, and explore the following data in one place.

Data type	Overview
Data Rows & Dataset	Data Rows in Catalog are imported when a dataset is created or appended with Data Rows.
Custom metadata	Custom metadata fields are imported during Data Row creation or update events.
Media attributes	Media attributes are a special class of metadata automatically pre-computed by Labelbox at Data Row creation or update events. The media attributes include file type, dimensions, and pre-computed embeddings. Media attributes are essential for your optimal experience with Labelbox.
Ground truth annotations	Ground truth annotations are available in the Catalog when ground truth annotations are created using Overview: Annotate data or imported Import annotations.
Model predictions	Model predictions are not yet supported in Catalog.

Data outflow

Data can flow out of the Catalog in two ways.

Type	Overview
Batch	Create a batch of Data Rows and send it to Annotate for labeling. Create a batch for labeling
Export	Use Python SDK to retrieve Data Row content (asset URL, media attributes, metadata). Export Data Rows from dataset Export labels

Type

Overview

Batch

Create a batch of Data Rows and send it to Annotate for labeling.

Create a batch for labeling

Export

Use Python SDK to retrieve Data Row content (asset URL, media attributes, metadata).

Export Data Rows from dataset

Export labels

Annotation previews

While all ground truth can be viewed in the Editor, the Catalog additionally supports previewing ground truth annotations to accelerate deriving insights that lead to actions. Below are supported and planned annotation previews of data types.

Data Types	Annotation Types
Image	Supported
Text, Video, Geospatial tiled imagery, Documents, DICOM	Coming soon

Getting started with Catalog

Chose data configuration: IAM Integration or Signed URLs
Import data
Curate data
Create a batch