Auto Labeling Guide#

Bootstrapping Datasets with Auto Labeling

Level: Intermediate | Estimated Time: 30-45 minutes | Tags: Auto-Labeling, Annotation, Delegated Operations, Model Inference

This guide walks you through using FiftyOne’s Auto Labeling to rapidly bootstrap and refine labels on your dataset. You’ll learn how to:

  • Generate high-quality auto labels using state-of-the-art models

  • Analyze and review predictions with confidence-based filtering

  • Refine your labels with visualization tools (patches view, embeddings)

  • Systematically approve correct predictions and flag issues

  • Complete the auto labeling workflow to integrate labels into your dataset

Guide Overview#

Auto Labeling combines model inference with human verification to dramatically accelerate dataset labeling. The workflow consists of the following steps:

  1. Gather Your Data and Infrastructure - Prepare your dataset in FiftyOne and configure GPU orchestration

  2. Configure Auto Labeling Run - Configure and launch an auto labeling task and track progress

  3. Analyze Results - Review predictions and select samples for approval

  4. Visualize Embeddings - Generate patch embeddings and use it to analyze clusters of samples

  5. Finalize Workflow - Accept approved labels and discard problematic predictions

Prerequisites#

Note

Auto Labeling is available in FiftyOne Enterprise. If you are using open source FiftyOne and are interested in this feature, please reach out to Voxel51 sales.

Who Is This Guide For

This guide is designed for machine learning practitioners and data scientists who need to efficiently label large datasets. Whether you’re bootstrapping a new project or improving existing annotations, Auto Labeling provides a systematic approach to leveraging model predictions while maintaining human oversight.

Required Knowledge

  • Familiarity with the FiftyOne Enterprise App interface and basic operations

  • Understanding of your target task (detection, classification, segmentation)

  • Basic knowledge of machine learning models and confidence thresholds

System Requirements

  • FiftyOne Enterprise: This feature requires FiftyOne Enterprise with delegated operations

  • GPU Access: Orchestrator must have GPU resources for model inference

  • Storage: Sufficient object storage space for dataset media and label fields

Ready to Begin?#

Click Next to start with the first step: Prepare Your Dataset and Delegated Operators.