Syllabus Map

• Study map: Syllabus Study Map

Overview

• Augmentation increases data diversity without new labels.
• Goal: improve generalisation by making the model invariant to nuisance changes (position, lighting, style).
• Must preserve the label semantics for the target task.

Common Techniques

Geometric

• Horizontal flip (for natural images where left-right symmetry is valid).
• Random resized crop to change scale and viewpoint.
• Rotation and translation for small pose shifts.
• Affine/perspective transforms for camera distortions.

Photometric

• Brightness, contrast, saturation, hue jitter.
• Gaussian noise and blur for sensor/noise robustness.
• JPEG compression artifacts for real-world deployment robustness.

Occlusion and Mixing

• Random erasing / Cutout: mask random image regions.
• CutMix: paste region from image B into image A and mix labels.
• Mixup: convex combination of two images and labels.

• $\lambda \in [0,1]$ controls mixing strength.

Task-Specific Augmentation Rules

Classification

• Strong photometric + moderate geometric transforms usually help.
• Typical pipeline: random crop, flip, color jitter, normalization.

Object Detection

• Must transform boxes together with image.
• Use multi-scale resize, random crop with box-validity checks, horizontal flip.
• Avoid aggressive crops that remove all objects.

Segmentation

• Must transform masks with the same geometry as images.
• Use nearest-neighbor interpolation for mask resizing to avoid class bleeding.
• Photometric transforms apply only to image, not mask labels.

Policy-Based Augmentation

• AutoAugment: searches augmentation policies on a validation set.
• RandAugment: applies $N$ random transforms with global magnitude $M$.
• TrivialAugment: one random transform with random magnitude; simple and strong baseline.

Practical Pipeline Design

Step 1: Define label-safe transforms

• Keep only transformations that preserve task labels.
• Example: vertical flip is invalid for digit recognition but often valid for aerial imagery.

Step 2: Set magnitude ranges

• Start with conservative ranges and increase gradually.
• Too-strong augmentations can cause underfitting.

Step 3: Order transforms

• Typical order: geometric -> photometric -> normalization.
• Put stochastic regularizers (Cutout/Mixup/CutMix) near the end of pipeline.

Step 4: Validate with ablations

• Add one augmentation family at a time.
• Track validation accuracy/mAP/IoU to isolate impact.

Step 5: Match training and deployment

• Include corruptions expected at inference (blur, compression, low light).
• Do not train with unrealistic transforms that never appear in production.

Practical Notes

Augmentation Policy

• Keep augmentations label-preserving.
• Tune intensity based on task sensitivity.

Training Interaction

• Augmentation and regularization interact: stronger augmentation often allows larger models or fewer explicit regularizers.

Small-Data Regime

• For small datasets, augmentation can matter as much as architecture choice.