Hi, all
I’m curious about using data augmentation. I saw edge impulse docs upgrade accuracy of models.
But, when I use data augmentation option, the accuracy is dropped 2%.
ver 3(use data augmentation) project ID : 23379
The result of model test which doesn’t use data augmentation option is around 89%, but when using data augmentation option it drops 87%.
except the data augmentation all of the options are same.
Actually, 2% of accuracy drops are not that big effects. I’m just curious about this events.
Could you explain about this events or any example which uses data augmentation and improves accuracy.
Hi @SunBeenMoon, the biggest benefit data augmentation gives is that it allows you to train longer without overfitting, so typically with data augmentation enabled you want to up the number of training cycles a little bit. Also, just the accuracy number does not tell the full story. Augmented models might work better on real-world data, can better handle situations with background noise (if there’s none of that in your model), etc. That’s the hard part of neural networks, there’s a lot of information and behavior in them that’s hard to fit into a single number.
any example which uses data augmentation and improves accuracy.
Keyword spotting with data augmentation
Keyword spotting without data augmentation
2 Likes
I was impressed by your kind response
Thanks!!! 
2 Likes
For FOMO Vision Data Augmentation, is there any explanation of what the augmentation does only from the point of improving my bounding boxes? I am not sure if data augmentation works best with a tight fit, perfect fit or large bounding box showing more of the background? Anyone any opinions?
