Launch: Quickly Label Images with Few-Shot AI Prompting

We are happy to announce a new AI Labeling feature that utilizes few-shot learning and visual prompting to significantly speed up annotation tasks. When labeling images with many similar objects, this new feature lets you identify a few examples of the object and then instantly apply the same label to similar objects.

Saving time on annotation with automatic AI Labeling

Imagine you have hundreds of images containing similar objects, such as a collection of quality assurance photos of products on an assembly line – or perhaps photos of packaged goods loaded into shipping cartons. If your production process requires identifying and labeling recurring objects across those images, you might spend a significant amount of time ensuring each object is labeled correctly.

The new Box Prompting feature in Roboflow’s AI Labeling product is perfect for these situations. Instead of manually labeling each object one-by-one, you can draw a box around one object to instantly identify and apply the same label to similar objects.

“I’ve seen annotation tasks completed in half the time with this new AI Labeling feature.”
- Shantanu, Roboflow

Getting started with visual prompting for data labeling

Click the “Box Prompting (AI Labeling)” tool in the annotation toolbar, draw a bounding box over one of the items, select an object class, and hit “Predict.” After that, you can adjust the confidence interval, manually tune labels, and then accept the annotations.

Let’s take a look at a few examples where few-shot prompting can speed up the annotation process.

Identifying beverage cans

Here is an example of labeling one beverage can and quickly identifying the other cans in an image.

Labeling medical vials

Perhaps you’re developing an internal application that recognizes medical vials placed in a tray. Again, you can quickly train a model to predict and find all vials after annotating one example.

Identify defects when packing products

If you are packing fragile products like eggs, you might want to identify the number of intact or broken eggs in each carton, so your system can flag the carton for attention.

Real-time learning improves labeling accuracy

As you fine-tune and approve predictions using the few-shot labeling feature, you are creating a model that adapts and learns in real-time based on your feedback. Since the prediction model persists across images, you can quickly apply labels and reduce time spent processing large sets of images.

“Many engineers spend a lot of time reviewing data pipelines and ensuring everything is labeled correctly. We built this feature so your team can minimize the hours spent labeling data without sacrificing quality of your annotations.”
- Jacob, Roboflow

Availability

The latest AI Labeling tools are available now for all Roboflow users. Simply log into your Roboflow account, start annotating an image, and select Box Prompting in the toolbar.