Image Augmentation in Practice — Lessons from 10 Years of Training CV Models and Building Albumentations

TL;DR Image augmentation is usually explained as “flip, rotate, color jitter”. In practice it operates in two very different regimes: In-distribution augmentation – simulate variations your data collection process could realistically produce Out-of-distribution augmentation – deliberately unrealistic perturbations that act as regularization Both are useful — and many high-performing pipelines rely heavily on the second. This guide explains how to design augmentation policies that actually improve generalization, avoid silent label corruption, and debug failure modes in real systems. The ideas here come from roughly a decade of training computer vision models and building Albumentations (15k GitHub stars, ~130M downloads). Contents The intuition: transforms that preserve meaning Why augmentation helps: two regimes The one rule: label preservation Build your first policy: a starter pipeline Prevent silent label corruption: target synchronization Expand the policy deliberately: transform