This hands-on workshop explores computer vision techniques for automotive damage detection using the CarDD dataset - the largest public dataset specifically designed for vehicle damage analysis. Participants will learn how to leverage FiftyOne, a powerful computer vision experimentation platform, to explore, visualize, and build models for detecting six common types of vehicle damage: dents, scratches, cracks, glass shatter, tire flats, and broken lamps.

The workshop consists of five comprehensive modules that guide participants from initial setup to advanced model deployment:

Environment Setup: Configure a Python environment with essential libraries including FiftyOne, PyTorch, and other dependencies needed for computer vision tasks.

• Dataset Exploration: Load and analyze the CarDD dataset, which contains 4,000 high-resolution images with over 9,000 expertly annotated instances of vehicle damage. Explore dataset statistics, annotation formats (COCO), and understand the dataset's split into training, validation and test sets.
• Visual Embeddings Analysis: Implement state-of-the-art vision models (CLIP, SigLIP) to gain deeper understanding of the dataset. Visualize relationships between damage types using dimensionality reduction techniques and explore similarities through natural language search capabilities.
• Model Evaluation: Learn techniques for evaluating model performance on instance segmentation and detection tasks. Export data for training and evaluate models using FiftyOne's built-in evaluation tools.
• Extended Functionality: Enhance your workflow using the FiftyOne plugin ecosystem for specialized visualizations, search capabilities, and integration with other tools. Deploy pre-trained "zoo" models for real-world car damage detection applications.

This workshop is designed for computer vision practitioners, automotive industry professionals, and data scientists interested in applying modern deep learning techniques to vehicle damage detection. Participants will gain practical experience working with a production-grade dataset while learning best practices for model development, evaluation, and deployment.