Label Validation Basics

Explorer Access by Role

The Explore tab is available on Workflow projects only. ManualQA projects do not provide Explore tab access for any role.

Users with the Viewer role on a Workflow project can now open the Explore tab and browse the Explorer in read-only mode. Viewers can access the grid, analytics, and embeddings views and apply existing filters, but Encord hides all mutation affordances for Viewers.

The following actions are not available to Viewers:

• Bulk actions (priority, collections, bulk classify, change class, delete labels, export CSV)
• Creating, editing, or deleting collections
• Adding or removing items from collections (collection tags remain visible)
• Creating, editing, deleting, or saving filter presets (including the Alt+P keyboard shortcut)
• Importing or deleting prediction sets (Manage button is hidden)

Viewers can post comments and raise issues in the Explorer, consistent with platform-wide permissions.

For the full role permissions matrix, see Roles and Permissions.

Compute Metrics and Embeddings

Computing metrics and embeddings is required to use Similarity Search, Natural Language Search, Quality Metric filters, and Embeddings Analysis.

1. Navigate to Projects.
2. Open the Explore tab.
3. Hover over Metrics & Embeddings
4. Click Compute.

5. Configure the computation:

   • Select whether you want Similarity & Natural language search and quality metrics
   • Select which embeddings to use.
   • Select whether you want the embeddings plot, and Diversity and Uniqueness metrics.

6. Click Start computation.

Quality Metrics

Quality metrics are only calculated when you compute metrics and embeddings.

Quality metrics evaluate your data, labels, and model predictions, forming the foundation of effective data curation. They provide meaningful ways to surface, rank, and explore your data — helping you identify issues, spot patterns, and make informed decisions about what to curate, fix, or prioritize.

Video Quality Metrics: Video quality metrics must be calculated by upgrading your folder. Examples include Area, Clip duration, Frames per second, Number of frames.

Data Quality Metrics: Data quality metrics must be calculated by upgrading your folder. Examples include Area, Frame number, Random value.

Data Quality Metrics

For more detailed information on Data Quality Metrics, refer to the Data Quality Metrics documentation.

Title	Metric Type	Ontology Type
Area - Ranks images by their area (width/height).	image
Aspect Ratio - Ranks images by their aspect ratio (width/height).	image
Blue Value - Ranks images by how blue the average value of the image is.	image
Brightness - Ranks images by their brightness.	image
Contrast - Ranks images by their contrast.	image
Diversity - Forms clusters based on the Ontology and ranks images from easy samples to annotate to hard samples to annotate.	image
Frame Number - Selects images based on a specified range.	image
Green Value - Ranks images by how green the average value of the image is.	image
Height - Ranks images by the height of the image.	image
Object Count - Counts number of objects in the image.	image	bounding box, checklist, point, polygon, polyline, radio, rotatable bounding box, skeleton, text
Object Density - Computes the percentage of image area that is occupied by objects.	image	bounding box, polygon, rotatable bounding box
Randomize Images - Assigns a random value between 0 and 1 to images.	image
Red Value - Ranks images by how red the average value of the image is.	image
Sharpness - Ranks images by their sharpness.	image
Uniqueness - Finds duplicate and near-duplicate images.	image
Width - Ranks images by the width of the image.	image

Label Quality Metrics

Label Quality Metrics are used for sorting data, filtering data, and data analytics.

Title	Metric Type	Ontology Type
Absolute Area - Computes object size in amount of pixels.	image	bounding box, polygon, rotatable bounding box
Aspect Ratio - Computes aspect ratios of objects.	image	bounding box, polygon, rotatable bounding box
Blue Value - Ranks annotated objects by how blue the average value of the object is.	image	bounding box, polygon, rotatable bounding box
Border Proximity - Ranks annotations by how close they are to image borders.	image	bounding box, point, polygon, polyline, rotatable bounding box, skeleton
Brightness - Ranks annotated objects by their brightness.	image	bounding box, polygon, rotatable bounding box
Broken Object Tracks - Identifies broken object tracks based on object overlaps.	sequence, video	bounding box, polygon, rotatable bounding box
Classification Quality - Compares image classifications against similar images.	image	radio
Confidence - The confidence that an object was annotated correctly.	image	bounding box, polygon, rotatable bounding box
Contrast - Ranks annotated objects by their contrast.	image	bounding box, polygon, rotatable bounding box
Green Value - Ranks annotated objects by how green the average value of the object is.	image	bounding box, polygon, rotatable bounding box
Height - Ranks annotated objects by the height of the object.	image	bounding box, polygon, rotatable bounding box
Inconsistent Object Class - Looks for overlapping objects with different classes across frames.	sequence, video	bounding box, polygon, rotatable bounding box
Inconsistent Track ID - Looks for overlapping objects with different track IDs across frames.	sequence, video	bounding box, polygon, rotatable bounding box
Label Duplicates - Ranks labels by how likely they are to represent the same object.	image	bounding box, polygon, rotatable bounding box
Missing Objects - Identifies missing objects based on object overlaps.	sequence, video	bounding box, polygon, rotatable bounding box
Object Classification Quality - Compares object annotations against similar image crops.	image	bounding box, polygon, rotatable bounding box
Occlusion Risk - Tracks objects and detects outliers in videos.	sequence, video	bounding box, rotatable bounding box
Polygon Shape Anomaly - Calculates potential outliers by polygon shape.	image	polygon
Randomize Objects - Assigns a random value between 0 and 1 to objects.	image	bounding box, polygon, rotatable bounding box
Red Value - Ranks annotated objects by how red the average value of the object is.	image	bounding box, polygon, rotatable bounding box
Relative Area - Computes object size as a percentage of total image size.	image	bounding box, polygon, rotatable bounding box
Sharpness - Ranks annotated objects by their sharpness.	image	bounding box, polygon, rotatable bounding box
Width - Ranks annotated objects by the width of the object.	image	bounding box, polygon, rotatable bounding box

Video Quality Metrics

Title	Metric Type	Ontology Type
Area - Ranks videos by their area (width/height).	video
Aspect Ratio - Ranks videos by their aspect ratio (width/height).	video
Blue Value - Ranks videos by how blue the average value of the video is.	video
Brightness - Ranks videos by their brightness.	video
Clip Duration - Ranks videos based on the video’s duration.	video
Contrast - Ranks videos by their contrast.	video
Diversity - Forms clusters based on the ontology and ranks videos from easy samples to annotate to hard samples to annotate.	video
Frame Number - Selects videos based on a specified range.	video
Frame Label Count	video
Frames Per Second	video
Green Value - Ranks videos by how green the average value of the video is.	video
Height - Ranks videos by the height of the video.	video
Instance Label Count - Ranks videos by the number of unique objects in the video.	video	bounding box, checklist, point, polygon, polyline, radio, rotatable bounding box, skeleton, text
Red Value - Ranks videos by how red the average value of the video is.	video
Sharpness - Ranks videos by their sharpness.	video
Uniqueness - Finds duplicate and near-duplicate videos.	video
Unlabelled Frames (%) - Ranks videos based on the percentage of unlabelled frames in the video.	video
Unlabelled Frames (#) - Ranks videos based on the number of unlabelled frames in the video.	video
Width - Ranks videos by the width of the video.	video

Model Quality Metrics

Model quality metrics help you evaluate your data and labels based on a trained model and imported model predictions.

Acquisition Functions

Acquisition functions are a special type of model quality metric, primarily used in active learning to score data samples according to how informative they are for the model, enabling smart labeling of unannotated data.

Title	Metric Type	Data Type
Entropy - Ranks images by their entropy.	image
Least Confidence - Ranks images by their least confidence score.	image
Margin - Ranks images by their margin score.	image
Variance - Ranks images by their variance.	image
Mean Object Score - Ranks images by their average object score.	image	object

Collections

Collections are saved groups of data units or labels that let you curate subsets of your data and perform bulk actions on them, such as sending items to annotation, running bulk classifications, or exporting a curated Dataset.

In the Explorer filter toolbar, each collection has a … (three-dot) menu that displays the collection name and its unique collection ID. Click the copy icon next to the ID to copy it to your clipboard — useful for API calls, support requests, or cross-referencing collections. Actions in this menu are organized into logical groups (navigate, manage, data, workflow, and danger) separated by dividers, making it faster to locate the right action. Empty groups are hidden automatically so no stray dividers appear.

Analytics View

Use the Analytics view to display Metric Correlation and prediction distribution for your ML model. Prediction distribution provides class and underrepresented class data. You can adjust the X and Y values on the Metric Correlation across a number of data and label metrics.

You can create custom analytics dashboards from the Analytics View using Distribution and Correlation charts. The Analytics View includes a Slice strip at the top of the canvas that lets you define named filter sets called slices and compare data distributions across multiple cohorts simultaneously.

Cohort Comparison via Slices

Encord lets you define named filter sets called slices and compare data distributions across multiple cohorts simultaneously on the same chart. Each slice appears as its own series, enabling side-by-side comparison without manually switching filters.

The Slice Strip

A Slice strip appears at the top of the Analytics canvas and is the primary control for creating and managing slices. It contains:

• All data — an unfiltered baseline chip that is always present. You cannot remove or rename it.
• Current filters — a transient chip that appears automatically whenever you have active toolbar filters. It shares the same filter set as the grid view, and has a dashed border to distinguish it from saved slices. Clicking it opens the Current filters editor.
• User-defined slices — named slices you create by clicking + Add slice.
• Add new chart — opens a menu to add a Distribution or Correlation chart to the canvas. This button replaces the former floating bottom-right button.
• How it works? — opens an illustrated modal explaining bucket-width snapping, shared-grid rebinning, and how hiding slices affects charts.

Slices are saved to your browser’s local storage, keyed by project. They persist across page reloads but are specific to your browser and are not shared with other collaborators.

Adding and Managing Slices

1. Add a slice

   Click + Add slice in the Slice strip. Encord creates a new empty slice and immediately opens the slice editor popover.

2. Configure the slice

   In the slice editor, you can:

   • Rename the slice.
   • Choose a color for the slice’s series on charts.
   • Set filters using the filter controls and any saved presets to define the cohort.
   • Click View in Explorer to apply the slice’s filters as the active filter set and navigate to the grid view.
3. Save the slice

   Close the slice editor popover. The slice chip appears in the Slice strip and Encord immediately adds it as a series on all distribution charts.

4. Edit a slice

   Click the pencil icon on a slice chip to reopen its editor and change its name, color, or filters.

5. Hide a slice from charts

   Click the eye icon on a slice chip to hide that slice. Encord excludes it from chart queries and the comparison footer entirely, not just from the visual display. Click the icon again to restore it.

6. Remove a slice

   Click the remove icon on a slice chip to delete it permanently.

Multi-Series Distribution Charts

When you have saved slices or an active Current filters set, distribution charts render as multi-series grouped bar charts — one bar series per visible slice. The All data baseline always appears as a series unless you hide it. A comparison footer below each chart mirrors the visible series from the Slice strip with a per-cohort count summary.

Each distribution chart also offers two bucket alignment modes, which you can toggle directly on the chart:

Mode	Description
Aligned	Rebins all series onto a shared bucket grid so bars align side by side for direct comparison.
Native	Keeps each series’ own backend bucket edges as translucent overlapping bars on a linear axis, preserving each cohort’s natural distribution shape.

Empty Analytics Dashboard

When the Analytics canvas has no charts, Encord displays a headline and quick-start cards. Click a card to add a Distribution or Correlation chart directly to the canvas.

The following do not support multi-series slice comparison and continue to use a single-series chart:

• File and frame client metadata distributions (some metadata endpoints)
• Feature-hash attribute distributions
• Text-search boolean distributions
• Prediction mode distributions
• Correlation charts (scatter plots), which always display data for the current unfiltered view

Distribution charts

Distribution charts display distributions and summaries of the selected metrics and custom metadata.

Here are some examples:

• Data unit: Frame number, random value, area
• Metadata: Enum with their enum options, numeric, date time, and boolean

varchar (previously string), text (previously long_string), and uuid are NOT SUPPORTED for use in Distribution charts.

When slices are present or active filters exist, Distribution charts render as multi-series grouped bar charts (see Cohort Comparison via Slices above for details on slices and bucket alignment modes).

Correlation charts

Correlation charts display a scatter plot of two attributes to show correlation within your current filtered view. Correlation charts require numeric data.

Distribution charts support a number of custom metadata types, however Correlation charts ONLY SUPPORT numeric custom metadata.

Correlation charts always display data for the current unfiltered view and do not support multi-series slice comparison.

Stat tiles

Stat tiles display a horizontal bar chart with one bar per slice. Each bar’s length is proportional to the item count of the largest slice currently in view, so you can compare slice sizes at a glance even when a small slice sits alongside a much larger baseline. The right-hand label shows the item count and the percentage of the baseline count.

Bar order runs from top to bottom: the unfiltered All data baseline appears first, followed by any active toolbar filter slice, then your user-defined slices.

When an embedding-distance filter is active, Stat tiles cannot retrieve a count for the affected slice and displays a dash (-) in place of the count.

Stat tiles defaults to small (S) size on the dashboard canvas. Encord saves your chart configuration across reloads.

Metadata coverage

Metadata coverage displays one row per metadata key. Each row contains one horizontal bar per slice, showing how many items in that slice have that key populated. All bars share a common scale — the highest single-slice count across all displayed rows — so you can compare coverage both across slices within a row and across rows for the same slice.

Use this chart to identify which metadata fields carry signal (high coverage) and which are sparsely filled across your dataset slices.

Selection modes

Metadata coverage offers two ways to choose which keys to display:

• Auto mode (default) — Encord automatically shows the top 20 keys by total coverage across all slices.
• Explicit mode — You select specific keys using the Keys multi-select dropdown. Selecting or removing any key switches the chart into explicit mode.

To remove an individual key from the chart, click the × button on that key’s row. To return to auto mode, click the Reset to top N link in the chart caption.

Metadata coverage supports both file-level and frame-level metadata domains.

If your project has no client metadata schema, or if the data query fails, Metadata coverage displays clearly labeled example data so the chart remains usable. The chart caption indicates when example data is showing.

Metadata coverage defaults to medium (M) size on the dashboard canvas. Encord saves your chart configuration — including any explicit key selections — across reloads.

Project Overview summary cards

The Overview tab of a workflow Project displays a set of summary cards that give you an at-a-glance view of Project health and progress.

All summary cards described below are hidden for labeling-services-only Projects.

Stat cards

A row of three compact stat cards appears at the top of the overview:

• Tasks completed — shows completed tasks out of the total task count, with a Project status selector in the footer.
• Time spent — shows the total time annotators have spent on the Project. When time data is available, a Show breakdown link navigates to the Collaborators tab in Analytics.
• Opened issues — shows the number of currently open (unresolved) issues. When open issues exist, a Show tasks link navigates to the Project queue filtered to tasks with unresolved issues.

Class distribution card

The Class distribution card shows a horizontal bar chart of label instances by class. It displays the top 5 and bottom 5 classes by instance count. When more classes exist between those two groups, the card shows a truncation indicator for the hidden middle classes.

• The card subtitle shows the total number of label instances (for example, “1,234 label instances”).
• When no label instances exist yet, the card displays “Instance labels by class will appear here.”
• When data is available, a Show analytics link navigates to the Labels tab in Analytics. This link is hidden when the card is empty.

Top issue tags card

The Top issue tags card replaces the previous Issues overview card. It shows up to 10 issue tags ranked by frequency, displayed as amber-colored horizontal bars. The card subtitle shows the total number of issue tags created for the Project.

• An info popover on the card explains what issue tags are and links to documentation.
• A Show analytics link navigates to the Issues tab in Analytics. This link is hidden when the card is empty.
• When no issue tags have been used yet, the card displays “Most common issue tags will appear here” and a Setup issue tags link to issue-tag settings.

Tag bar counts reflect tag-action totals rather than unique issue counts. The count for untagged issues is a frontend approximation and may under-report — it can show 0 even when untagged issues exist.

Custom Analytics Dashboard

Before creating a custom analytics dashboard, we recommend having custom metadata available in your data. Custom metadata can make the insights you get from the dashboard much more useful.

1. Navigate to Project > Explore and select a folder.

2. Click Analytics view.

3. Specify the display criteria for the Distribution and Correlation cards that display by default.

4. Click Add chart to add additional Distribution and Correlation cards.

5. Specify the display criteria for the added Distribution and Correlation.

Crop View

Crop View is only available from the Labels page of the Explore tab in your Project.

The Labels page displays all object and classification annotations on your images and video frames. Crop View zooms in on each annotated object, making it easier to inspect small or densely annotated regions.

For example, if a blueberry is annotated in an HD video frame containing many blueberries, Crop View off shows the full frame with the annotation highlighted but hard to see. Crop View on zooms directly into the annotated object.

To turn ON Crop View for all labels:

1. Navigate to your Project and click Explore.

2. Click Labels. The Labels page appears.

3. Click Display. The Display tab appears.

   

4. Toggle the Crop View switch. Object labels immediately are zoomed in on. Images/video frames with Classifications remain unchanged.

   Classifications on images/video frames are not affected by the Crop View feature. This is because Classifications apply to the entire image/video frame, while object annotations apply to specific areas/regions of an image/video frame. The following image has a Classification label/annotation Blueberry or Cherry? Blueberry and a bitmask object label/annotation Blueberry. The bitmask object annotation zooms in, while the classification does not.

5. Adjust the Crop View Zoom as required.