Retail shrink from shoplifting — curated open dataset + synthetic augmentation

Executive Summary

Retailers collectively lose billions to shrink each year, and shoplifting—both opportunistic and organized—remains a primary driver; loss-prevention teams at roughly 120,000 enterprise locations are already budgeting for analytics and software to reduce those losses. Current computer-vision approaches are hampered by severe data scarcity for rare events, inconsistent labeling across sites, and privacy/legal barriers that make it hard to build robust, production-grade detectors. You could build a curated, open video dataset of annotated shoplifting and loss-related incidents paired with a synthetic-augmentation pipeline that generates diverse, rare-event variants (different camera angles, clothing, occlusion, group dynamics) plus standardized labels and evaluation suites. Package this with baseline on-device models, SDKs for inference, and a subscription/licensing model so retailers, system integrators, and researchers can train, benchmark, and deploy solutions faster. The timing is attractive: about 120,000 enterprise retailers represent a ~$12B addressable market at roughly $100k ACV per customer, edge compute costs have dropped enough to enable low-latency on-device inference, and recent advances in transformers, self‑supervised learning, and synthetic-data tooling materially lower the cost to create diverse, realistic training examples. To stand out, prioritize high-quality curation, rigorous anonymization and compliance workflows, realistic physics- and behavior-based synthesis, and an independent benchmark that reduces vendor lock-in; these make the dataset practical and defensible. Be honest about the challenges: obtaining consented, legally compliant real-world footage, dealing with domain shift between store formats and cameras, and competing with established security vendors who may develop or acquire similar assets.