Exploring the Voxel51 FiftyOne Computer Vision Toolkit as a Centralized Platform for Cross-Disciplinary Data Analysis, Modeling, Embedding and Visualization
This repository contains the collective work of Group 18 (2025) for the Autumn Advance Data Science Internship conducted by the IDEAS – Institute of Data Engineering, Analytics and Science Foundation, ISI Kolkata.
The project demonstrates the versatility of the Voxel51 FiftyOne toolkit as a centralized platform for data analysis, modeling, embedding, and visualization across diverse scientific domains.
FiftyOne is a powerful open-source toolkit designed to enhance the quality of datasets and computer vision models. It provides a flexible and interactive environment to visualize, curate, and analyze complex data, bridging the gap between raw data and machine learning models.
This repository is structured into five chapters, each exploring a unique application of the FiftyOne toolkit:
- Chapter 1: Medical Image Analysis of IDRiD Dataset - Aliasgar Saria
- Analysis of the Indian Diabetic Retinopathy Image Dataset (IDRiD), demonstrating dataset management, disease severity prediction (96% accuracy), and the use of FiftyOne's "Brain" for advanced error analysis.
- Chapter 2: Astronomical Morphology Classification - Sk Salman Parbhage
- A novel multi-representation ensemble learning methodology on the Galaxy10 DECals dataset, achieving 99.66% classification accuracy and using FiftyOne for comprehensive misclassification analysis.
- Chapter 3: Image Deduplication - Mukesh G
- A systematic workflow for data cleaning using perceptual hashing to efficiently identify and manage duplicate images, showcasing FiftyOne's capabilities in data curation.
- Chapter 4: Physical Sciences Data Analysis - Arja Banerjee
- A novel application of FiftyOne to a numerical dataset of specific heat capacities, successfully visualizing and validating the classical Dulong–Petit Law.
- Chapter 5: Foundational Computer Vision Exploration - Sritoma Roy
- A core demonstration on the Caltech101 dataset, showing how to generate and visualize image embeddings to uncover semantic structures, clusters, and anomalies.
- Primary Framework: Voxel51 FiftyOne
- Programming Language: Python
- Machine Learning: Scikit-learn, XGBoost, PyTorch
- Key Concepts: Deep Learning Embeddings (ViT, EfficientNetV2, CLIP), Perceptual Hashing, Dimensionality Reduction (UMAP, t-SNE)
- Interactive Data Visualization: Using the FiftyOne App to explore datasets, visualize complex labels, and analyze model predictions.
Documentation and Presentations: (https://tinyurl.com/y66nx2y5)
- Advanced Analytics with FiftyOne Brain: Leveraging tools for similarity searches, uniqueness detection, and identifying labeling mistakes.
- Comprehensive Dataset Management: Efficiently organizing, querying, and manipulating large and diverse datasets.
- In-depth Model Evaluation: Analyzing model performance through interactive confusion matrices and precision-recall curves.
This project serves as a practical guide and a testament to FiftyOne's capability as a unifying platform for data-centric AI workflows across various scientific and technical fields.