How to Speed Up Image Labeling: Toolbox Features That Save Clicks

To speed up image labeling, cut the clicks per object and reuse work: pick the cheapest shape the model needs, pre-label with a model, clone objects between similar images, and use hotkeys. The Supervisely image annotation tool puts all of these in one browser-based labeling toolbox.

Why is image labeling slow, and how much time can you save?

Drawing is the bottleneck. A study of crowdsourced box labeling reports that the classic way of drawing a box took 35 seconds per high-quality box on ILSVRC, while extreme clicking needs about 7 seconds. Shape matters even more. The Cityscapes team spent more than 1.5 hours per image on fine polygon annotation and under 7 minutes per image on coarse polygons.

Scale makes it worse. The COCO dataset has 2.5 million labeled instances in 328k images. Most real projects are smaller, but a few thousand images with ten objects each still means tens of thousands of shapes.

Source	What was measured	Number
Extreme clicking, Papadopoulos et al.	Time per box, extreme clicking	about 7 s
Same paper	Time per box, classic drawing on ILSVRC	35 s
Cityscapes	Fine polygon annotation and quality control, per image	more than 1.5 h
Cityscapes	Coarse polygon annotation, per image	under 7 min
MS COCO	Labeled instances in the dataset	2.5 million in 328k images

The rest of this guide covers what to change in your workflow, and which Supervisely features do the work.

What does a fast image labeling workflow look like?

Four levers give most of the gain. The table shows which toolbox feature pulls each one. Everything listed is described in the Supervisely docs for the image labeling toolbox.

Lever	Toolbox feature	Use it when
Cheaper shape	Bounding Box	The model predicts boxes (object detection)
Fewer clicks per object	Auto-select, hotkeys, crosshair	Many objects per image
Model does the first pass	NN Image Labeling app, Smart Tool	You have a model or a similar dataset
Reuse between images	Clone objects and tags, multi-view	Frames or views are close to each other

If you are comparing tools rather than workflows, see the labeling toolbox page and the feature comparison.

What is new in the Supervisely image labeling toolbox?

The toolbox was redesigned in 2023-2024. Here is what a labeler gets:

1. Web-based interface: you only need a browser.
2. Customizable layout and visualization: light and dark theme, flexible layout, multiple image view modes, multi-spectral view and grid.
3. Complex image formats: high-resolution images, high-color depth images with 16 bits per pixel or more, image metadata, restore mode and undo/redo.
4. Several labeling tools: Bounding Box, Polygon, Brush and Eraser, Mask Pen, Smart Tool and Graph (Keypoints), with 1000+ objects per image, image and object tags, and customizable hotkeys.
5. Teamwork features for large annotation teams, and integration with neural networks and AI-assisted tools.
6. Import and export in many formats, including medical NRRD, NIfTI and DICOM data.

You can try it for free in the Community Edition, no credit card needed.

How to set up the workspace for speed?

Start with the layout. Every panel can be dragged, resized and moved, so you can build a workspace that fits your task.

Layout components of the new Image Labeling Toolbox

The numbers in the screenshot:

1. Basic settings: history of operations, theme, hotkeys map and more.
2. Working area for the image and its labels.
3. Objects panel with classes and tags.
4. List of images in the dataset, apps you can embed into the toolbox, visualization and other settings.
5. Annotation tools.

Basic interface elements

1. Undo reverts the last action, so you can try a different approach without fear.
2. Redo brings back an undone action.
3. Theme: switch between light and dark. Handy for long shifts 🦉☕.
4. Hotkeys: the map of shortcuts that speed up labeling.
5. Full-screen mode maximizes the working area.
6. Screenshot: capture the visible area and share it with colleagues.
7. Restore mode brings back deleted objects ever created in the project.
8. Multi-view image mode shows several images on one scene.

Which hotkeys matter most?

Keyboard shortcuts switch tools, move objects and change visualization faster than any menu. They are customizable, so you can keep the combinations you know. The toolbar also shows a tooltip with a short GIF and the hotkey for each tool.

Customizable hotkeys available in the New Image Labeling Toolbox for enhanced workflow

A few from the docs: SHIFT+H randomizes object colors, SHIFT plus the mouse wheel changes opacity, and the right arrow key clones selected objects to the next image.

Which annotation tool should you use?

Pick by what the model has to predict, not by what looks best. Even if a neural network pre-labels your data, you still need manual tools to correct its mistakes and to build test and validation sets.

Supervisely complete set of image annotation tools

Task	Shape	Tool	Guide
Object detection	Rectangle	Bounding Box	Bounding box annotation
Segmentation with clear edges	Polygon	Polygon	Polygon annotation
Irregular or thin shapes	Mask	Brush and Eraser	Brush tool
Masks with both straight and free edges	Mask	Mask Pen	Mask Pen tool
Fast masks from a few clicks	Mask	Smart Tool	Smart Tool
Poses and skeletons	Keypoints	Graph	Human pose estimation

Bounding Box

Bounding boxes are rectangles with a class, used in object detection to find and localize objects. They are defined by coordinates that give the position and size of the object.

Detection models predict boxes, so training data can be exported in all popular annotation formats. Read the full bounding box annotation guide.

Polygon Tool

Polygon outlines object boundaries by connecting points. Unlike boxes, polygons can have any number of points, so they fit irregular shapes. The tool supports quick creation, editing of single points and polygon parts, snapping to existing object boundaries and automatic point placement.

See polygon annotation best practices for semantic and instance segmentation.

Mask Pen Tool

Mask Pen merges polygon and brush features: switch between modes, add or remove mask parts, fill holes with one click, or split a mask into two objects. Native mask support also lets you upload segmentation model predictions to the platform for review.

Brush Tool

Brush creates free-form masks. It is optimized for huge resolutions and works with hundreds of masks on one scene. It can pick scattered parts of an object into one mask, fill holes with one click, and use the Polygon split mode to split masks.

How to label faster with AI assistance?

Let a model draw the first version and spend your time on corrections. There are two ways in the toolbox.

Smart Tool for masks

Smart Tool is an interactive segmentation tool. A class-agnostic model produces a mask from your clicks. You add positive and negative points to fix it. It works with models such as RITM, ClickSeg and Segment Anything. You can also train a custom Smart Tool on your own data without code. See AI-assisted labeling for the full picture.

NN Image Labeling for any model

The NN Image Labeling app connects a neural network to the labeling interface. Models like YOLOv8 and OWL-ViT predict classes and tags, which are added to the project automatically.

Supervisely App

NN Image Labeling

Use deployed neural network in labeling interface

Run in SuperviselyGitHub

How it works:

• Any model with a serving app, such as Serve MMDetection 3.0, can be connected.
• Add the app to your team from the Ecosystem and deploy the model in the team first. The app does not support multiuser mode, so each user runs their own session, but several sessions can connect to one model.
• Since version 1.6.0 you can add a suffix to model class and tag names to avoid conflicts with existing ones.
• The app can apply the model to a region of interest defined by an object bounding box.

Want to train your own detector first? Read the no-code YOLOv8 tutorial. For labeling with one example, see OWL-ViT one-shot detection.

Which productivity features save the most clicks?

Auto-select objects

When Auto-select is on, hovering over an object selects it if it has the current shape. This removes a click per object and helps most when an image has many objects.

Clone objects and tags

The Objects panel copies selected objects to the next image, or brings them from the previous one. The Images panel does the same for tags. This is a big saving for sequences and near-duplicate frames.

Multi-image view

Multi-view mode shows several images on one scene, grouped by tag value. The project needs a tag with values like "Text" or "One of", and the images need to carry it. Use it to label grouped images together, compare results, or label dependent images such as medical studies (NRRD, DICOM, NIfTI), objects captured by several cameras or geospatial data. More in multi-view image annotation.

Grid for large images

The NxM grid sets a comfortable cell size and guides you through the areas of a high-resolution image with many small objects.

Grid view of planting a small tree garden

Tags and properties

Tags describe images and objects: classification, search, training. A tag can hold a string, a number, a one-of value, or no value at all. Assign a hotkey, a color, and choose if the tag fits images, objects or both. Details in the image and object tags guide.

Filter images and objects

Combine conditions to find exactly the images or objects you need in a large dataset, for example to review one class or to find unlabeled images.

Customizing the image filters

Customizing the object filters

Metadata

You can preview and edit image metadata and custom data. Custom data is useful for storing integration information when you connect Supervisely with internal systems.

Image metadata always available

How to adjust the scene so you can see small objects?

All visualization settings are in one panel, so you spend less time in menus.

Hovering over each function displays a brief description of its functionality

• Opacity: hold and drag left or right. SHIFT plus the mouse wheel changes it from anywhere on screen.
• Border: drag to change border width. Useful for huge resolutions or many small objects.
• Point: drag to change the radius of object points.

• Default color paints objects with their class color.
• Randomize color (click, then SHIFT+H) separates objects of the same class, which helps in instance segmentation.

Attribute display settings show ID, bindings (see objects binding), tags, classes and the creator near objects. The visibility mode can be Always, Show on hover or Show when selected.

Image display settings adjust brightness and contrast for dark or low-contrast images, and let you zoom in to label object boundaries precisely.

Objects and images panels

The Objects panel manages tags, attributes and metadata. You can filter objects, hide all but one class, merge objects, delete them or change their layering. The Images panel lists the dataset, lets you filter and edit metadata, download images or annotations, and delete images.

The Apps panel opens apps from the Computer Vision Ecosystem inside the toolbox. Python developers can write custom apps that react to labeling events in real time, so a labeler sees them as built-in features. The Settings panel personalizes the interface. See custom labeling UI.

How to import data and export labels?

Ecosystem apps import images in 50+ annotation formats from local storage, other platforms and cloud services. Supervisely detects the annotation format by default. Browse the import apps.

Application catalog for Import available in the Supervisely Ecosystem Apps

The Auto Import app creates a project and adds images in these formats: .jpg, .jpeg, .mpo, .bmp, .png, .webp, .tiff, .tif, .nrrd, .jfif, .avif, .heic, NIfTI, DICOM.

Export apps move annotations, images and other data to the formats that neural network frameworks expect. Browse the export apps.

Application catalog for Export available in the Supervisely Ecosystem Apps

How to speed up labeling with a team?

Split the dataset between labelers with labeling jobs or labeling queues. A job defines the task, the object types, the guidelines and the access: a labeler sees only the datasets and classes assigned to them. You track progress and review finished work. Teams that need to run the platform on their own servers can use the on-premise Enterprise Edition, and the security page lists the compliance details. See also user collaboration.

You can also take a screenshot of the working area and share the link with a colleague, without leaving the platform.

How to check quality while you label?

The integrated Dataset Ninja statistics app shows class distribution, a confusion matrix, image and object properties, heatmaps, treemaps and spatial heatmaps for your own dataset in a single click, and keeps them updated as annotations change. More on quality assurance.

Integrated Dataset Ninja's dynamic heatmap charts, treemaps, and spatial heatmaps for immersive statistics analysis and visualization

Is the Supervisely labeling toolbox right for you?

Supervisely suits teams that need several shapes (boxes, polygons, masks, keypoints), model-assisted labeling and team workflows in one place, with cloud or on-premise deployment. If you only need boxes on a handful of images, a lighter tool may be enough. Check the feature comparison and the pricing page.

See all features of the Supervisely image annotation tool or read the image labeling toolbox docs.