Visual QC Review

Visual QC (Quality Control) lets you review model predictions by comparing them against ground truth labels. This process helps measure how much manual correction would be needed in real-world usage and identifies common prediction errors.

Navigation Flow

Visual QC follows a three-level navigation:

1. Tags — Select a tag category to review
2. Models — Select a specific model (training run)
3. Review — Review batch images and rate predictions

Step 1: Select a Tag

The main Visual QC page shows all available tags as cards.

Each card displays:

• Tag name and status
• Pattern used for matching
• Number of models reviewed and average VQC score (if available)

Step 2: Select a Model

After selecting a tag, you'll see a table of all models available for review.

Column	Description
Model ID	Unique identifier for the training run
Training Date	When the model was trained
Test mAP50	Model accuracy from automated testing
VQC Score	Manual review score (calculated from your ratings)
Review Status	Shows whether review is not started, in progress, or completed
Last Review	Date of most recent review activity

Click a row to open the review interface for that model.

Step 3: Review Images

The review interface displays batch images in a side-by-side comparison view with a clickable grid overlay for rating individual images.

Interface Layout

Area	Function
Header	Shows overall review progress (images reviewed across all batches) and activity log button
Presence Banner	Appears when other reviewers are viewing the same model (see Collaboration)
Left Panel	Labels (Ground Truth) — what the model should detect
Right Panel	Predictions (Model Output) — what the model actually detected
Grid Overlay	Configurable grid (2×2 to 5×5) dividing the image into cells — click any cell to rate it
Navigation Bar	Previous/Next buttons, batch selector dropdown with status indicators, grid size selector, and zoom controls
Status Bar	Numbered chips showing which images have been reviewed, plus "Mark Remaining as Blank" action

Navigation Controls

• Arrow Buttons — Move to the previous or next batch
• Batch Dropdown — Jump to any batch directly. Each batch shows a status indicator:
  • Green "Done" chip — all images in the batch have been reviewed
  • Blue "3/16" chip — shows how many images are reviewed out of total
  • No chip — batch has not been started
• Keyboard Shortcuts — Use Left/Right arrow keys to navigate between batches

Tip: Your current batch position is saved in the URL. If you refresh the page or share the link, it will return to the same batch.

Grid Overlay

The grid overlay divides the batch image into individual cells. Each cell represents one image in the batch and can be reviewed independently.

• Grid Size — Select 2×2, 3×3, 4×4, or 5×5 from the dropdown in the top-right corner. Choose the size that matches your batch mosaic layout. Your preference is saved per model.
• Cell States — Each cell shows its review status visually:
  • Unreviewed — Subtle green highlight on hover, shows cell number
  • Reviewed — Green background with a checkmark icon
  • Blank — Dark overlay with strikethrough text
• Click any cell to open the rating modal for that image.

Handling Batches with Different Image Counts

Different batches within the same model may contain different numbers of real images. This is normal — it typically happens because:

• The last batch of a model often contains fewer images than earlier batches if the total number of test images isn't evenly divisible by the batch size
• Some models were trained with a different mosaic layout (e.g., 2×2 instead of 4×4), so their batch composites have fewer cells

You might see this as:

• Black or empty-looking cells at the end of a batch
• A grid that looks "larger than the actual content" — for example, a 4×4 grid with 16 cells but only 1–4 real images in the top-left

How to handle it:

1. Rate each real image as you normally would (Good, Minimal Intervention, Poor, etc.)
2. For cells that are clearly empty or all-black, select "Blank / Empty Cell" as the assessment
3. Or use "Mark Remaining as Blank" (see below) to handle all empty cells in the batch with one click after rating the real images

Note: The grid size you select applies to the entire model, not to individual batches. If most of your batches are 4×4 but one batch only has a few images, that's expected — use the Blank rating to handle the empty cells. Do not change the grid size to match one odd batch, as it will affect all batches.

Zoom and Pan

Inspect details by zooming into the images:

• Zoom In/Out — Click the +/− buttons or use your scroll wheel (up to 8× magnification)
• Pan — Click and drag when zoomed in to move around
• Reset — Click the reset button or double-click the image to return to default view

Tip: Zoom is synchronized between the Labels and Predictions panels. When you zoom or pan in one panel, the other follows automatically, making it easy to compare the same region.

Rating Images

Click any grid cell or numbered chip in the status bar to open the rating modal.

Overall Assessment (Required)

Select one assessment that best describes the prediction quality:

Assessment	Description
No Intervention	Perfect prediction, no correction needed
Minimal Intervention	Minor adjustments needed (1-2 clicks)
Moderate Intervention	Some corrections needed (3-5 clicks)
Major Intervention	Significant corrections needed (6+ clicks)
Not Usable	Prediction is completely wrong, must redo
Not Relevant	Image not applicable for this class
Blank / Empty Cell	No image in this cell (padding in the batch mosaic)

Error Tags (Optional)

Select up to 2 tags to identify specific issues with the prediction:

• AI Extraneous — False positive (detected something that shouldn't be there)
• AI Missed — False negative (missed something that should be detected)
• AI Wrong Class — Detected the right area but classified incorrectly
• Label Error — The ground truth label is incorrect
• Poor Image Quality — Image is blurry, dark, or hard to analyze
• Other — Other issue not covered above

Notes (Optional)

Add any observations or comments about the image for future reference.

Mark Remaining as Blank

If a batch has empty cells — either at the end of the mosaic (padding) or because the batch has fewer images than the grid suggests — you can quickly mark all unreviewed cells as "Blank" in one action instead of rating them individually. See Handling Batches with Different Image Counts for common reasons this happens.

1. Review at least one image in the batch first.
2. Click "Mark Remaining as Blank" in the status bar at the bottom.
3. Confirm in the dialog that appears.

Note: This button only appears when there are unreviewed images in the current batch, and requires at least one image to already be reviewed. You can check "Don't show this again" to skip the confirmation for the rest of your session.

Activity Log

The activity log shows a timeline of all reviews for the current model. Click the history icon in the header to open it.

• All tab — Shows reviews across every batch in the model
• Current Batch tab — Filters to only show reviews for the batch you're currently viewing

Each entry shows the assessment given, the batch and image number, who performed the review, and when it happened. The log loads more entries automatically as you scroll down.

Collaboration Features

Visual QC supports multiple reviewers working on the same model simultaneously. When other users are viewing the same model, a presence banner appears below the header.

The banner shows:

• Viewer avatars — Colored circles with initials for each active reviewer
• Batch location — Which batch each reviewer is currently viewing
• Same batch warning — If another reviewer is on the same batch as you, a yellow "same batch" chip and "Changes may conflict" message appear

Note: To avoid conflicts, coordinate with team members when reviewing the same batch simultaneously. Consider splitting work by batch ranges.

Progress Tracking

Track your review progress using these indicators:

• Header Progress Bar — Shows total images reviewed across all batches with a percentage
• Batch Dropdown Chips — Green "Done" chip on fully reviewed batches, blue progress count on partially reviewed batches
• Grid Overlay — Green cells with checkmarks for reviewed images, dark cells for blank images
• Status Bar Chips — Numbered chips at the bottom: green (reviewed), strikethrough (blank), outlined (unreviewed)

Review Tips

• Compare the same region in Labels vs Predictions to spot differences — zoom is synchronized across both panels
• Use zoom to inspect small or unclear detections
• Use error tags to identify patterns in prediction errors
• Add notes for unusual cases that might need team discussion
• Use Left/Right arrow keys for faster batch navigation
• Use "Mark Remaining as Blank" to quickly handle batches with empty padding cells — especially useful for the last batch, which often has fewer real images than earlier batches
• Set your grid size to match the batch mosaic layout — it's saved automatically per model
• Check the activity log to see what others have reviewed and where work is still needed