OCAIS internship project
Dataset Quantifier
Software Engineer Intern
Built Python tools for image-quality analysis, AI/ML dataset evaluation, and image-processing performance investigation.
THE BRIEF
What needed to be built.
Understand differences across image datasets and identify images or model runs that need closer investigation.
MY CONTRIBUTIONS
My contribution.
Dataset preparation
- Processed annotated and unannotated image datasets and generated CSV, JSON, and PDF outputs.
- Implemented image rotation and training/test dataset preparation.
- Built capture-ingestion endpoints that validate image filenames and annotation metadata, preserving the association between uploaded images and CSV/JSON analysis records.
Analysis & performance
- Built a Python multiprocessing runner that partitions image folders across GPU-bound workers and skips folders already marked as processed.
- Implemented filtering and clustering to analyze image characteristics and model evaluation scores.
- Instrumented model runtime and grouped outliers to identify unusually slow image processing and support debugging.
THE RESULT
Image processing reduced from 3s to 0.1s per image.
Reduced image processing from 3s to 0.1s per image during my OCAIS internship. Built dataset-analysis and visualization tools for Mask2Former and YOLOv11L evaluation, contributing to a 20% improvement in model accuracy.