label-quality

The toolkit to test, validate, and evaluate your models and surface, curate, and prioritize the most valuable data for labeling.

Consistency and Accuracy analysis on CelebA

A Python toolkit for cleaner datasets in computer vision.

(WIP): 'Aporia' in Greek means 'inconsistent'. A Python library that detects and fixes dataset issues using both rule-based methods and ML models. It evaluates dataset quality across multiple metrics, including missing values, duplicates, outliers, class imbalance, and label consistency. It also suggests fixes based on the metric scores.

Machine learning-based quality assurance of object bounding-box labels: code for the Applied Sciences 2023 paper and the 2025 PhD thesis, plus a unified implementation

Production-grade annotation QA & inter-annotator agreement pipeline with from-scratch Cohen/Fleiss/Krippendorff metrics, tiered review workflows, and guideline revision diagnostics.

Web image scraper for object detection, to be extended to other tasks and modalities: collects image URLs from Common Crawl pages, detects Pascal VOC objects and scores label quality; searchable and exportable behind a subscription (Flask API, React app, Chrome extension, offline pipeline)

Rule checkers ask "is there a label?". Signpost asks "does the label mean anything?" — a small multilingual model for accessibility label quality, plus a constrained rewriter.