Increase image-classification training data.
Do not change the label meaning.
Choose two practices.
Apply only transformations that preserve labels to training data and keep validation and test data independent real examples.
Detailed explanation
Only meaning-preserving transformations are used.
Evaluation data leaks.
Artificially inflated evaluation is avoided.
Training labels become wrong.
Check official documentation
Compare the following points with the official documentation.
Review Task 1.3 and the preprocessing and splitting guidance. For hypothetical flipped or cropped defect images, distinguish label-preserving from label-changing transformations and prevent an original image and its derivatives from crossing split boundaries.
Official references
What to confirm
Distinguish increasing training examples from independent evaluation, preserving label meaning and separation at the original-data level.
Key points
- Augmentation
- Label preservation
- Independent evaluation
Notes
- Environment: Reference review of official documentation and hypothetical examples; no AWS resources or model training are required.
- Putting derivatives of the same source image in training and test data undermines evaluation on unseen images. Split with source-image or same-subject groups in mind.
- Additional transformed evaluation sets can be used for robustness checks. Preserve independent source data and distinguish the ordinary evaluation from these additional checks.
Foundation review
Data augmentation
Creating additional training examples by applying controlled transformations to existing data.
Label-preserving transform
A transformation that changes presentation without changing the correct target label.