Annotate Data

Annotations teach your VLM what to look for and how to respond. Use Datature's visual annotation tools to create high-quality training data for phrase grounding and visual question answering tasks.

📋

Prerequisites

• A dataset with uploaded images
• Understanding of your use case and annotation requirements

Create a dataset if you don't have one yet.

---

What is annotation?

Annotation is the process of adding labels to your images to teach your model what to recognize and understand. For vision-language models, annotations connect visual information (images) with text descriptions, questions, or instructions.

Why annotations matter:

• Training data — Annotations are examples your model learns from
• Task definition — What you annotate determines what your model can do
• Model accuracy — High-quality annotations lead to better predictions
• Use case alignment — Annotations should match your real-world application

Think of annotations as teaching by example—the more accurate and comprehensive your annotations, the better your model will understand and perform.

---

Annotation types

Datature supports two main annotation types for vision-language models. Choose based on your use case and what you want your model to do.

Phrase Grounding

Phrase grounding connects natural language descriptions to specific regions in an image. You write a caption describing the image, then link phrases to bounding boxes around the objects they describe.

Use cases:

• Object detection with natural language ("Find the large black chip")
• Visual search and retrieval
• Image description and captioning
• Zero-shot object detection

What you create:

• Image captions describing visible objects
• Bounding boxes around objects
• Links between phrases and boxes

Learn how to annotate for phrase grounding →

Visual Question Answering (VQA)

Visual Question Answering teaches your model to answer questions about images. You create question-answer pairs that help the model understand visual content and make decisions.

Use cases:

• Quality control and inspection ("Is this product defective?")
• Inventory and counting ("How many items are on the shelf?")
• Compliance monitoring ("Is safety equipment worn?")
• Condition assessment ("What is the crop health status?")

What you create:

• Questions about image content
• Clear, concise answers
• Multiple question types (counting, yes/no, attributes, categories)

Learn how to annotate for VQA →

---

Choose your annotation approach

Select the annotation type that best fits your use case. You can use both types in different datasets for different applications.

---

Annotation workflow

Follow this general workflow to create high-quality annotations efficiently.

1. Prepare your dataset

Before annotating, ensure your dataset is ready:

• Upload images to your dataset
• Review image quality and coverage
• Define your annotation goals and requirements
• Create annotation guidelines for consistency

Learn about dataset management →

2. Choose your annotation type

Select based on your use case:

• Phrase Grounding — Link text to objects
• VQA — Create question-answer pairs

3. Annotate your images

Use Datature's visual annotation tools:

• Open the annotator from your dataset
• Create annotations systematically
• Use AI-assisted tools to speed up annotation
• Maintain consistency across your dataset

4. Review and refine

Ensure annotation quality:

• View dataset insights to analyze coverage
• Review annotations for accuracy and consistency
• Edit or remove low-quality annotations
• Add more annotations if needed

5. Train your model

Use your annotations to fine-tune a VLM:

• Create a training workflow
• Configure your training settings
• Start a training run
• Evaluate model performance

---

AI-assisted annotation

Speed up annotation with AI-powered tools that suggest captions, phrases, and annotations automatically.

💡

IntelliScribe features

Datature's AI-assisted annotation tools help you annotate faster while maintaining quality:

• Auto-caption generation — Generate image descriptions automatically
• Phrase highlighting — Automatically link phrases to bounding boxes
• Smart suggestions — AI recommendations based on image content

Learn about AI-assisted tools →

Benefits of AI assistance:

• Speed — Annotate 3-5x faster with AI suggestions
• Consistency — AI maintains consistent terminology
• Quality — Review and refine AI suggestions for accuracy
• Scalability — Annotate large datasets efficiently

Best practice: Use AI assistance to accelerate annotation, then review and refine suggestions manually for optimal quality.

---

Annotation best practices

Follow these guidelines to create high-quality training data that improves model performance.

Consistency

• Use standardized terminology across your dataset
• Follow the same annotation patterns for similar images
• Create and follow annotation guidelines
• Review annotations regularly for consistency

Quality over quantity

• Accurate annotations are more valuable than many low-quality ones
• Take time to annotate carefully and thoughtfully
• Review and edit annotations when you spot errors
• Focus on clear, unambiguous annotations

Coverage and diversity

• Annotate objects at different scales and positions
• Include various lighting conditions and angles
• Cover edge cases and challenging scenarios
• Balance your annotation distribution across categories

Efficient workflow

• Learn keyboard shortcuts for faster annotation
• Use AI-assisted tools appropriately
• Annotate similar images in batches
• Take breaks to maintain annotation quality
• Track your progress regularly

---

Collaborative annotation

Work with your team to annotate large datasets efficiently.

Team annotation workflow:

1. Add team members to your organization
2. Create shared annotation guidelines based on documentation
3. Assign different images or batches to different annotators
4. Review annotations for consistency across team members
5. Use dataset insights to track team progress

Tips for team consistency:

• Reference the detailed annotation guides: Phrase Grounding or VQA
• Create a style guide with examples specific to your use case
• Schedule regular review sessions to maintain quality
• Use consistent terminology and answer formats
• Share best practices and learnings within your team

---

What's next?

✅

Ready to annotate

Choose your annotation type and start creating high-quality training data for your vision-language models.

Start annotating

---

Related resources

• Upload images — Add images to your dataset before annotating
• Create a dataset — Set up a dataset for your annotation project
• Manage datasets — Organize and maintain your annotated datasets
• Train a model — Use your annotations to fine-tune a VLM
• Phrase grounding concepts — Deep dive into phrase grounding
• Visual question answering concepts — Deep dive into VQA
• Add team members — Collaborate on annotation projects
• Annotate for phrase grounding — Step-by-step phrase grounding guide
• Annotate for VQA — Step-by-step VQA annotation guide
• AI-assisted tools — Speed up annotation with IntelliScribe
• Upload annotations — Import existing annotations
• Download data — Export datasets and annotations
• View dataset insights — Check annotation progress and quality
• Quickstart — End-to-end training workflow