Computer vision & evaluation
Visual robustness
A change of scenery.
Blur it. Add noise. Change the light. Compare what three training recipes help a vision model handle.
Try a transformation
Turn it up45%
An illustration of the transformation, using the portfolio’s editorial image. No classifier runs here. The trained models’ measured accuracy is compared below.
01 / The question
Does the recipe survive a different-looking image?
02 / What I built
I trained ResNet-18 from scratch on 100,000 Tiny ImageNet images for 100 epochs, comparing baseline augmentation, RandAugment, and AugMix under a shared setup.
I evaluated ten corruption types at five severity levels: 50 controlled visual-shift conditions. Clean accuracy was kept alongside corruption accuracy so the trade-off stayed visible.
03 / What happened
AugMix raised mean corruption accuracy from 27.47% to 35.32%, about 7.9 percentage points. Clean accuracy was 61.64% versus the baseline’s 61.66%. RandAugment also improved corruption accuracy while retaining clean performance.
| Training recipe | Clean | Corrupted |
|---|---|---|
| Baseline | 61.66% | 27.47% |
| RandAugment | 62.14% | 34.40% |
| AugMix | 61.64% | 35.32% |
04 / Where it stops
This is a controlled study of one architecture and dataset. Synthetic corruptions do not cover every real distribution shift. The interactive image above illustrates a transformation; it does not run the trained classifiers or predict their accuracy at the selected setting.
05 / Keep exploring
One more? / 06