Problem Images

The Problem Images page tracks Visual QC assessments to identify training images that require attention. It aggregates assessment data across model versions to help you prioritize which images need correction or removal.

Page Overview

The page displays:

• Assessment summary - Counts of images by assessment level
• Performance metrics - Average mAP50, Precision, and Recall
• Filters - Filter by tag and assessment type
• Images table - Expandable rows with per-model details

Assessment Levels

Images are categorized by the level of intervention needed:

Assessment	Meaning
No Intervention	Prediction is correct, no action needed
Minimal Intervention	Minor adjustments needed (1-2 corrections)
Moderate Intervention	Some corrections needed (3-5 corrections)
Major Intervention	Significant corrections needed (6+ corrections)
Not Usable	Prediction is completely wrong, cannot be used
Not Relevant	Image not applicable for this class

Filtering Images

Use the filters to focus on specific problem areas.

Tag Filter

Select a specific tag from the dropdown to see only images for that category.

Assessment Filter

Click the assessment chips to filter by intervention level. You can select multiple assessments. Selected chips appear solid colored. Click Clear to reset the filter.

Images Table

The table shows all problem images with sortable columns:

Column	Description
Image Path	File path of the image (hover to see full path)
Tag	Category the image belongs to
Assessment	Overall assessment level (worst across all models)
Models	Number of models this image was assessed in

Recurring Issues

Images that appear in multiple models are marked with a RECURRING badge.

Recurring issues indicate that the same image causes problems across different model versions. These should be prioritized for correction as they consistently affect model performance.

Viewing Assessment Details

Click any row to expand it and see per-model assessment details.

The expanded view shows:

• Model ID - Which model the assessment came from
• Assessment - The intervention level for this specific model
• Errors - Error tags identifying specific issues (e.g., AI Missed, AI Extraneous)
• Notes - Any additional notes from the reviewer

Error Tags

Error tags help identify the specific type of prediction error:

• 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

Exporting Data

Click the Export CSV button to download all problem images matching your current filters.

The exported CSV includes:

• Image Path
• Tag
• Overall Assessment
• Models Count
• Model Details (Model ID, Assessment, Errors, Notes for each model)

Tip: Use the filters to narrow down the export to specific tags or assessment levels before exporting.

Tips for Using Problem Images

• Focus on Major Intervention and Not Usable images first
• Check for RECURRING issues that affect multiple models
• Look at error tags to understand why predictions failed
• Use the Label Error tag to identify ground truth issues that need correction
• Export data to share with the annotation team for review