2D Object Grounding

Precise object grounding with bounding boxes and points

2D Object Grounding

Qwen3-VL provides precise 2D object grounding capabilities, allowing you to locate and label objects within images using relative position coordinates. The model supports both bounding boxes and point-based grounding for diverse positioning and labeling tasks.

Capabilities

Qwen3-VL’s 2D grounding uses relative position coordinates to:

• Bounding Boxes: Draw rectangular boxes around objects
• Point Grounding: Mark specific locations with coordinate points
• Flexible Combinations: Mix boxes and points for complex annotation tasks
• Multi-object Detection: Ground multiple objects simultaneously
• Relative Coordinates: Position-independent coordinate system (0-1 range)

How It Works

Coordinate System

Positions are expressed as relative coordinates:

• X-axis: 0 (left edge) to 1 (right edge)
• Y-axis: 0 (top edge) to 1 (bottom edge)

This makes grounding resolution-independent and portable across different image sizes.

Grounding Formats

1. Bounding Boxes: [x_min, y_min, x_max, y_max]
2. Points: [x, y]
3. Combined: Mix both formats in a single response

Use Cases

• Object Detection: Locate and label objects in images
• Annotation: Generate training data for computer vision models
• Visual Search: Find specific items within images
• Quality Control: Identify defects or anomalies in products
• Spatial Analysis: Analyze object positions and distributions
• Interactive Applications: Enable click-to-identify features

Try It Out

Explore 2D object grounding with our interactive cookbook:

2D Grounding Cookbook

Using relative position coordinates, it supports both boxes and points, allowing for diverse combinations of positioning and labeling tasks.

Key Features

• High Precision: Accurate object localization
• Format Flexibility: Support for boxes, points, and combinations
• Multi-object Support: Ground multiple items in one pass
• Resolution Independent: Works across different image sizes
• Natural Language Queries: Describe what to ground in plain text

Advanced Capabilities

Referring Expression Comprehension

Ground objects based on natural language descriptions:

• “The red car on the left”
• “The person wearing glasses”
• “The largest apple in the bowl”

Dense Captioning

Generate descriptions for multiple grounded regions in an image.

Visual Question Answering with Grounding

Answer questions while providing visual evidence through grounding.

Related Capabilities

• 3D Grounding - 3D bounding boxes for spatial scenes
• Omni Recognition - Identify what objects to ground
• Spatial Understanding - Understand spatial relationships
• Video Understanding - Grounding in video frames