Choosing Augmentations: Hypotheses, Protocols, and Metrics

Creating a new augmentation pipeline is rarely a straightforward task. It often begins with basic crops and reflections, gradually incorporating elements from previous projects, articles, and competitions. Eventually, numerous transformations accumulate without a clear rationale for their selection. This article presents a structured method for approaching this challenge: understanding the purpose of each transformation, what it simulates, how intensely it should be applied, and the assumptions it makes about the data. The central idea is that augmentation serves as a clear statement about which variations should not alter the meaning of the label. This clarity aids in deciding what to retain or discard and helps distinguish genuine benefits from mere noise in the training process. There is no 'magic solution'; instead, the focus is on intuition, a mental model, and a step-by-step protocol for practical systems. The article includes an engineering perspective, two levels of analysis, a seven-step pipeline, strength and budget adjustments, advanced techniques, diagnostics, metrics, indicators of harm, automated searching, deployment, and examples.