Gaussian Noise Augmentation

Additive Gaussian Noise Augmentation

Upon request we investigated if additive Gaussian Noise as form of data augmentation could yield any benefits for our baseline.

During training each patch is augmented with a new random sample of additive Gaussian Noise with zero-mean and standard deviation of 0.2. The same noise is applied to the validation data. During testing we look at both options (i) with noise augmentation during testing (train & test) and (ii) without noise augmentation during testing (train only).

Method	Average Precision
Baseline	0.578 +/-0.01
Baseline with Noise Augmentation (train & test)	0.566
Baseline with Noise Augmentation (train only)	0.597
DenoiSeg	0.625 +/-0.008
DenoiSeg with Noise Augmentation (train & test)	0.674
DenoiSeg with Noise Augmentation (train only)	0.607

Since the baseline and DenoiSeg are run five times we additionally report one standard error about the mean. For the noise augmentation experiments only a single run is currently available.

With this small set of experiments we can conclude that noise augmentation can help in general, but has a more profound beneficial effect for DenoiSeg, giving additional merit to our proposed training approach.