Skip to content
BENIBIX Learn

Visual Grounding

AI🎨 MultimodalVision and Language

Also known as Phrase Grounding, Referring Expression Grounding, Grounded Image Understanding

Finding the spot in a picture that some words point to. A box marks it.

Think of it like

It is like asking a friend where the red cup is. She puts her finger on it.

Example

Jack's app is told to find the blue coat by the door. In a photo of a busy hall, it draws a box around that one coat.

Why it matters

Many tasks need more than knowing what is in a picture. They need to know where the named thing is, so they can act on it.

Definition

Visual grounding is finding the exact part of a picture that some words describe. You give the model an image and a phrase, like 'the red cup on the left'. It answers with a box or an outline around that thing. This is harder than spotting cups in general, because the words pick out one of them. It lets a robot, an app or an AI agent point at things it was told about in plain words.
Middle: the words name one cup, and the model draws a box around just that one.

Common mistake

Mixing it up with finding objects. That finds every cup. Grounding finds the one cup the words point to.

Key takeaways

  1. 1
    Words to a place. A phrase becomes a box.
  2. 2
    One, not all. The words pick out a single thing.
  3. 3
    Lets AI point. Useful for robots and agents.

Compare