Computer Vision Dataset Quality Auditor

Computer Vision Dataset Quality Auditor helps machine-learning engineers, computer-vision teams, AI startups, annotation vendors, MLOps teams, researchers, and dataset curators identify hidden training-data weaknesses before expensive model training, evaluation, fine-tuning, benchmarking, or production deployment.

The skill is designed for:

Machine-Learning Engineers Computer-Vision Engineers AI Startups Annotation Teams Data-Labeling Vendors MLOps Teams Data Scientists Applied Researchers Dataset Curators Quality-Assurance Teams Robotics Teams Industrial Inspection Teams Manufacturing AI Teams Retail Computer-Vision Teams OCR Teams Document-AI Teams Video-Analytics Teams Geospatial-Vision Teams Agriculture-Vision Teams Ecommerce Visual-Search Teams Multimodal-AI Teams Synthetic-Data Teams Responsible-AI Teams Data-Governance Teams

It can audit datasets used for:

Image Classification Multilabel Classification Object Detection Instance Segmentation Semantic Segmentation Panoptic Segmentation

OCR

Document Understanding Pose Estimation Keypoint Detection Object Tracking Video Classification Action Recognition Event Detection Visual Inspection Anomaly Detection Visual Search Image Retrieval Scene Understanding Robotics Perception Industrial Vision Multimodal AI

The agent performs a structured workflow:

Dataset Intake → Task Definition → Dataset Inventory → File Integrity Audit → Exact Duplicate Detection → Near-Duplicate Analysis → Split Leakage Review → Video Redundancy Analysis → Class Distribution Analysis → Annotation Quality Audit → Image and Video Quality Audit → Diversity Analysis → Bias and Shortcut-Learning Review → Deployment-Match Analysis → Severity Ranking → Remediation Planning → Re-Audit Criteria

The audit begins by establishing the intended computer-vision task and deployment conditions.

It identifies:

Dataset name Dataset version Business objective Vision task Target classes Deployment environment Expected cameras Expected devices Expected image resolution Expected object sizes Expected lighting Expected weather Expected geography Expected backgrounds Expected frame rates Expected capture distances Expected operating conditions Cost of false positives Cost of false negatives Safety relevance Privacy constraints

Dataset quality is evaluated relative to the intended task.

For example:

An image may have sufficient resolution for scene classification but insufficient detail for OCR.

Motion blur may be undesirable for product photography but representative and necessary for vehicle tracking.

Nighttime footage may appear visually difficult but be essential if the deployed system operates at night.

The agent creates a complete dataset inventory.

The inventory can include:

Total files Total images Total videos Total frames Number of classes Number of annotations Training samples Validation samples Test samples Source systems Capture devices Geographic regions Collection periods Annotation vendors Annotation batches Synthetic-data share Augmented-data share Known limitations

For each media item, the agent can track:

File identifier Relative path Media type Image or video format Width Height Aspect ratio File size Duration Frame rate Frame count Codec Color space Bit depth Orientation Dataset split Data source Class Annotation presence Corruption status Capture metadata Annotation version

The skill performs file-integrity analysis.

It can detect or organize findings involving:

Corrupted images Corrupted videos Zero-byte files Unreadable files Truncated media Decoder failures Unsupported formats Invalid dimensions Inconsistent file extensions Missing annotations Orphan annotations Duplicate filenames Path collisions Missing video frames Invalid timestamps Impossible frame rates Malformed annotation records Missing segmentation masks Invalid references

Each integrity finding can include:

Issue Affected count Affected percentage Affected files Severity Expected model impact Recommended action Verification requirement

The agent performs exact duplicate analysis.

Exact duplicates may represent:

Accidental duplicate ingestion Renamed copies Repeated exports Repeated reference images Duplicate frames Cross-split leakage

It can group duplicates into clusters and report:

Cluster identifier Files Classes Sources Dataset splits Duplicate type Leakage risk Recommended disposition

The skill also performs near-duplicate analysis.

Possible similarity techniques include:

Perceptual hashing Image embeddings Feature similarity Structural similarity Local-feature matching Crop-aware similarity Temporal similarity Video fingerprinting

Near duplicates may include:

Resized images Cropped images Recompressed images Brightness variants Contrast variants Watermark variants Minor edits Burst photographs Adjacent video frames Overlapping video clips Synthetic variants generated from nearly identical seeds

The agent does not automatically recommend removing every near duplicate.

It evaluates whether redundancy:

Inflates apparent dataset size Reduces effective diversity Creates train-validation leakage Represents legitimate deployment variation Supports a temporal task Provides meaningful viewpoint changes Is required by the evaluation design

The skill performs detailed train-validation-test leakage analysis.

Potential leakage patterns include:

Identical images appearing in multiple splits Near-identical images crossing splits Crops from the same original image crossing splits Saved augmentation derivatives crossing splits Adjacent video frames crossing splits Overlapping video segments crossing splits Images from the same capture session crossing splits The same physical instance appearing across splits when independence is required Before-and-after pairs split incorrectly Sequential inspection images distributed randomly Metadata exposing the target label Source-specific information correlated with labels

For video datasets, the agent recommends selecting an appropriate independence unit such as:

Video Clip Recording Session Event Sequence Subject Camera Session Location Day Production Batch

Random frame-level splitting is flagged when temporal dependence can inflate evaluation results.

The agent performs video redundancy analysis.

It can evaluate:

Repeated frames Near-identical adjacent frames Static scenes Redundant clips Overlapping windows Frame-sampling density Scene-transition frequency Positive-event density Negative-segment length Labeled-duration distribution Class duration imbalance Temporal annotation drift Event-boundary consistency

A video-redundancy finding can contain:

Video Segment Redundancy score Effective unique content Class Split Risk Recommended sampling strategy

The skill performs class-balance analysis appropriate to the task.

For classification, it can analyze:

Samples per class Unique sources per class Unique scenes per class Train-validation-test distribution Duplicate-adjusted sample counts

For object detection:

Images containing each class Instances per class Boxes per image Object co-occurrence Object size by class Rare-object distribution

For segmentation:

Images containing each class Instances Pixel share Object-area share Rare-region frequency

For video:

Clips Events Frames Labeled duration Unique sequences

Possible balance metrics include:

Class ratio Majority-to-minority ratio Relative class share Entropy Effective sample count Long-tail concentration Source-adjusted class count Unique-scene count

The skill does not assume equal class sizes are optimal.

Class representation is evaluated against:

Expected deployment prevalence Business consequences False-positive cost False-negative cost Rare-event importance Model objective Evaluation objectives Collection feasibility

The class-balance report can include:

Class Train count Validation count Test count Instance count Unique source count Relative share Expected deployment share Risk Recommended action

The skill performs annotation-taxonomy analysis.

It can identify:

Duplicate class names Synonyms used as separate classes Spelling variants Capitalization variants Deprecated labels Unused labels Overlapping class definitions Nonexclusive classes Hierarchy inconsistencies Label-map mismatch Annotation-guideline mismatch Class-definition drift

Label-quality findings are categorized as:

Confirmed Error Suspected Error Ambiguous Case Taxonomy Defect Annotation-Policy Ambiguity Annotator Inconsistency Missing Annotation Duplicate Annotation

The skill does not automatically relabel examples based solely on model predictions.

Model disagreement is treated as a review signal.

A labeling-risk record can include:

Sample Current label Suspected issue Supporting evidence Confidence Policy ambiguity Recommended human review Adjudication requirement

For object-detection datasets, the agent audits bounding boxes for:

Coordinates outside image bounds Zero-area boxes Negative dimensions Excessively loose boxes Excessively tight boxes Partially excluded targets Several objects merged into one box One object split incorrectly Incorrect classes Duplicate boxes Missing boxes Boxes covering background Truncation inconsistencies Occlusion inconsistencies Crowd-policy inconsistencies Tiny-object policy violations

Bounding-box analysis can include:

Normalized area Aspect ratio Distance to image edge Boxes per image Overlap distribution Class-specific object size Truncation frequency Occlusion frequency

For segmentation datasets, the skill can audit:

Empty masks Masks outside image bounds Invalid class indices Holes Disconnected fragments Boundary leakage Missing object regions Overlapping incompatible classes Self-intersecting polygons Mask-image misalignment Masks associated with the wrong image Low-resolution masks Bounding-box-to-mask inconsistencies

For pose and keypoint datasets, it can audit:

Keypoints outside image bounds Left-right swaps Incorrect body-part assignment Impossible skeleton geometry Missing visible keypoints Visibility-flag inconsistencies Duplicate person identifiers Occlusion-policy differences Annotation-format mismatch

For OCR and document-vision datasets, it can audit:

Incorrect transcriptions Bounding-region misalignment Missing characters Encoding problems Whitespace inconsistency Punctuation inconsistency Case normalization inconsistency Raw-text versus normalized-text confusion Unsupported scripts Wrong text orientation Language misclassification Duplicated documents Near-duplicate crops Synthetic-font imbalance Background bias Scan-quality imbalance

For tracking datasets, it can detect:

ID switches Duplicate track identifiers Missing track identifiers Track fragmentation Impossible jumps Unexpected class switches Visibility inconsistency Occlusion errors Incorrect start boundaries Incorrect end boundaries Cross-sequence leakage

The agent performs image-quality analysis.

Blur analysis distinguishes among:

Defocus Blur Motion Blur Intentional Depth-of-Field Blur Compression-Induced Softness

Possible measurements include:

Edge strength Laplacian variance Frequency-domain information Local sharpness Directional motion evidence

Blur is evaluated relative to deployment reality.

The agent does not automatically remove difficult images simply because they are blurry.

Resolution analysis can evaluate:

Image width Image height Megapixels Minimum object dimensions Target pixel area Text character height Evidence of upscaling Inconsistent resizing Class-specific resolution distributions

The agent assesses whether target features remain visible at the available resolution.

Compression analysis can identify:

JPEG blocking Ringing Repeated recompression Chroma artifacts Low video bitrate Macroblocking Frame corruption Encoding inconsistencies

Exposure analysis can identify:

Severe underexposure Severe overexposure Clipped highlights Clipped shadows Poor dynamic range

Contrast and color analysis can identify:

Very low contrast Extreme color cast Abnormal saturation Inconsistent white balance Class-specific color-processing differences

Orientation analysis can identify:

Incorrect rotation Upside-down images EXIF orientation mismatch Stretched images Unexpected aspect ratios Letterboxing Inconsistent crop policies

The skill evaluates target visibility.

It can analyze:

Occlusion Truncation Tiny objects Crowded scenes Overlapping objects Low foreground-background contrast Unusual viewpoints Unusual poses Reflections Transparency Glare Shadows Rain Fog Snow Night conditions Motion

These conditions are categorized as:

Valid Difficult Example Underrepresented Hard Case Mislabeled Difficulty Unusable Media Out-of-Scope Media

Valid hard examples should often be preserved because they improve real-world robustness.

The skill performs visual-diversity analysis.

Possible diversity dimensions include:

Lighting Time of Day Weather Season Geography Environment Indoor versus Outdoor Background Viewpoint Camera Angle Focal Length Zoom Distance Object Scale Object Pose Occlusion Truncation Color Texture Object Variant Camera Type Device Type Resolution Compression Frame Rate Motion Scene Complexity Contextual Co-Occurrence

Diversity is analyzed per class rather than only globally.

Each class can receive a coverage profile containing:

Lighting Coverage Viewpoint Coverage Scale Coverage Background Coverage Device Coverage Region Coverage Hard-Example Coverage Missing Conditions Risk

The skill performs visual-bias and shortcut-learning analysis.

Potential shortcut signals include:

Class strongly correlated with background Class strongly correlated with watermark Class strongly correlated with camera Class strongly correlated with image resolution Class strongly correlated with region Class strongly correlated with lighting Class strongly correlated with file format Class strongly correlated with annotation style Class strongly correlated with synthetic origin Positive and negative classes collected under different conditions Target object absent while context still predicts the class

Each shortcut-risk finding can include:

Potential Shortcut Affected Class Correlated Attribute Evidence Strength Alternative Explanation Model Risk Recommended Validation Remediation

Potential validation approaches include:

Background-only baseline Foreground-only baseline Source-held-out evaluation Stratified performance analysis Counterfactual evaluation Saliency inspection Attribution analysis

The skill does not claim that shortcut learning actually occurred without appropriate model-based evidence.

Source-bias analysis can compare data by:

Collection Source Camera Device Annotation Vendor Annotation Batch Annotator Group Region Contributor Data Partner Collection Campaign Capture Period Synthetic Generator Augmentation Pipeline

This can reveal:

Source-specific image quality Annotation drift Class-specific collection conditions Hidden source shortcuts Dataset-deployment mismatch

Background-bias analysis checks whether:

Positive examples consistently use distinct scenes Negative examples lack comparable backgrounds Synthetic examples contain recognizable rendering contexts Cropped data removes realistic environmental variation Scene context predicts class labels

Potential remediation can include:

Collecting hard negatives Adding counterexamples Diversifying backgrounds Collecting positives across additional environments Collecting negatives inside positive-class contexts Evaluating background-only performance

Camera and device bias can be evaluated across:

Sensor Manufacturer Lens Resolution Aspect Ratio Compression Color Profile Frame Rate Orientation Capture Software

The agent recognizes that models can learn device-specific artifacts instead of semantic target features.

Geographic and temporal diversity analysis can consider:

Country Region Climate Infrastructure Architecture Signage Vegetation Product Variants Collection Period Season Event Period Firmware Generation Product Generation

Sensitive demographic analysis is performed only when appropriate, authorized, governed, and supported by legitimate metadata.

The skill does not infer protected characteristics from visual appearance as a default dataset-audit technique.

The agent audits hard-negative coverage.

Hard negatives are examples visually similar to positive cases but not containing the target.

It evaluates:

Quantity Visual Similarity Source Diversity Background Diversity Environment Diversity Deployment Relevance Class Proximity

Insufficient hard-negative coverage can cause excessive false positives and misleadingly strong validation performance.

The skill audits synthetic data.

It can review:

Synthetic Share Class Distribution Generator Diversity Seed Diversity Background Diversity Lighting Realism Geometry Realism Texture Realism Boundary Artifacts Renderer Signatures Unrealistic Co-Occurrence Metadata Leakage Synthetic-to-Real Ratio Split Allocation Duplicated Generations Label Correctness Domain-Gap Risk

Synthetic data is not assumed to improve a dataset automatically.

The agent also audits augmentation pipelines.

It can review:

Augmentation Type Probability Class-Specific Application Geometric Validity Label Transformations Bounding-Box Transformations Mask Transformations Keypoint Transformations Unrealistic Transformations Saved Augmented Samples Crossing Splits

Augmentation should preserve the target semantics.

The skill analyzes metadata leakage.

Potential leakage sources include:

Class Names in Filenames Class-Specific Folder Paths Source IDs Timestamps Batch Names Watermarks Synthetic Markers Annotation-Tool Fields Borders Camera Serials Encoded Target Metadata

If metadata enters the model pipeline, target leakage can become critical.

The skill performs deployment-match analysis.

It compares the dataset with expected real-world conditions across:

Camera Resolution Lighting Object Distance Object Prevalence Background Region Device Compression Frame Rate Season Operating Environment Hard Cases User Behavior

A domain-gap record can include:

Dimension Training Distribution Expected Production Distribution Gap Severity Evidence Recommended Data Collection

The skill performs annotation-consistency and vendor-quality analysis.

It can compare quality by:

Annotator Vendor Batch Guideline Version Collection Period Class

Possible measures include:

Agreement Rate Class Agreement Bounding-Box IoU Mask Overlap Keypoint Distance Temporal-Boundary Agreement Adjudication Frequency

Annotation disagreement is not automatically treated as annotator failure.

It can indicate an unclear annotation policy.

The skill audits annotation guidelines for coverage of:

Class Boundaries Inclusion Rules Exclusion Rules Ambiguous Cases Occlusion Truncation Crowd Cases Tiny Objects Overlapping Objects Partial Objects Difficult Examples Ignore Regions Negative Examples Temporal Boundaries Keypoint Visibility OCR Normalization Quality Escalation Adjudication

When model predictions are supplied, the agent can use them to prioritize:

High-Confidence Disagreements High-Loss Samples Low-Margin Samples Embedding Outliers Repeated Failure Clusters Class-Cluster Inconsistencies

Predictions are treated as evidence for review rather than an automatic source of truth.

Embedding analysis can support:

Near-Duplicate Clustering Visual-Diversity Measurement Outlier Discovery Class-Overlap Analysis Source Clustering Synthetic-versus-Real Separation

An outlier can represent:

A mislabeled example A corrupted file A rare valid example A deployment-critical hard case A source anomaly An out-of-scope sample

Outliers are classified before removal.

The skill applies a severity framework.

Critical findings can include:

Major train-validation leakage Systemic label mapping errors Large-scale missing annotations Invalid dataset-split methodology Widespread corrupted annotations Sensitive-data exposure

High-severity findings can include:

Severe near-duplicate leakage Major rare-class undercoverage Strong annotation inconsistency Strong source bias Important deployment domain missing Large synthetic-to-real domain gap

Medium findings can include:

Moderate class imbalance Moderate blur concentration Uneven viewpoint coverage Source-specific quality drift Insufficient hard negatives

Low-severity findings can include:

Minor class-name inconsistencies Isolated low-resolution examples Low-impact metadata gaps

The skill can build component-level quality scores for:

File Integrity Duplicate Risk Split Integrity Class Coverage Annotation Accuracy Annotation Consistency Media Quality Visual Diversity Deployment Match Metadata Quality Shortcut-Learning Risk

An aggregate score must never hide critical weaknesses in individual dimensions.

The agent produces a remediation plan.

Possible remediation actions include:

Remove confirmed accidental duplicates Group related samples Rebuild dataset splits Relabel confirmed errors Adjudicate ambiguous examples Revise annotation guidelines Reannotate affected batches Retrain annotators Increase QC sampling Collect rare-class examples Collect hard negatives Collect missing environments Diversify cameras Diversify backgrounds Collect low-light examples Collect occluded examples Replace corrupted files Replace accidental low-resolution copies Normalize orientation Reduce video redundancy Resample temporal data Rebalance source contribution Validate synthetic examples Reduce excessive synthetic representation Correct taxonomy Improve provenance metadata

Each remediation item can include:

Issue Severity Affected Scope Recommended Action Expected Benefit Risk of Change Effort Priority Owner Validation Method

Remediation is prioritized using:

Evaluation Integrity Expected Model Impact Deployment Risk Class Importance Business Cost Safety Relevance Fairness Relevance Affected Sample Count Ease of Correction Collection Cost Labeling Cost Schedule Impact

The agent defines re-audit criteria.

A remediation should not be considered complete until:

Affected samples are corrected Dataset version is updated Split leakage is rechecked Class statistics are recalculated Media-quality metrics are rerun Annotation-quality checks are rerun Known issues are reassessed Remaining exceptions are documented Evidence is retained

The skill supports dataset-version governance.

Version records can include:

Dataset Version Parent Version Collection Additions Removed Samples Relabeled Samples Split Changes Taxonomy Changes Annotation-Policy Changes Synthetic-Data Changes Augmentation Changes Audit Date Reviewer Approval

Supported operating modes include:

Full Dataset Audit: Complete analysis of media, labels, splits, diversity, bias risk, and remediation.

Duplicate & Leakage Audit: Exact duplicates, near duplicates, temporal redundancy, related captures, and cross-split leakage.

Class Balance Audit: Class counts, rare classes, source-adjusted representation, and split balance.

Annotation Quality Audit: Taxonomy defects, suspected mislabels, boxes, masks, keypoints, OCR, tracking, and policy inconsistencies.

Image & Video Quality Audit: Blur, resolution, compression, exposure, orientation, corruption, and temporal redundancy.

Bias & Diversity Audit: Background, source, camera, geography, lighting, viewpoint, scale, and shortcut-learning risks.

Deployment Gap Audit: Compares training data with real-world production conditions.

Synthetic Data Audit: Evaluates synthetic realism, diversity, proportion, metadata leakage, and domain shift.

Annotation Vendor QA: Compares vendors, batches, annotators, and annotation-guideline versions.

Remediation Planner: Transforms findings into a prioritized, measurable correction backlog.

The core commercial promise is: transform a computer-vision dataset into a rigorous quality profile that reveals duplicate and leakage risk, class imbalance, annotation weaknesses, media-quality failures, missing diversity, potential visual shortcuts, deployment mismatches, and the exact remediation priorities needed to build more reliable models.