State the invariance assumption
An augmentation is useful when the transformed example should retain the intended label. Horizontal flips can be reasonable for some object categories but can change the meaning of text or directional signals. Specify which real-world variation the transformation represents instead of applying a standard menu blindly.
Explain why flipping a left-turn road sign should not automatically preserve its label.
Split originals before augmentation
Keep transformations of the same original in its training partition. Putting an original in training and a crop in evaluation leaks closely related information. Validation should primarily represent real deployment inputs; artificial stress tests can be useful but should be reported separately from ordinary accuracy.
Design a split for photographs of 200 physical objects, each photographed from five angles.
Run a bounded comparison
Hold architecture, training budget, split, and metric constant while comparing augmentation policies. Inspect transformed examples visually, including labels or bounding boxes. Strong cropping can remove the target entirely; color changes can erase a diagnostic signal. More variation is not always better.
Create three policies: none, mild brightness variation, and aggressive crops. State the evidence needed to select one.
Test the promised robustness
If augmentation is intended to improve low-light behavior, report performance on real low-light examples. Synthetic diversity does not establish coverage of a missing population or environment. Record the policy and randomization settings with the model, and ensure stochastic training transformations are not accidentally active during normal inference.
An augmented model improves validation but fails on a new camera. Explain what coverage evidence was missing.
Which transformation is safe by default?
Keep exploring
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.