0
$\begingroup$

I am training on yolo and I had a small dataset. I decided to increase it by augmenting it with rotation, shearing, etc to increase the size and increase accuracy.

Now I have seen augmented datasets labeled as with and without original images.

I was wondering if there is difference between training with and without original images besides there just being more images?

$\endgroup$
4
  • $\begingroup$ How many image for class in your dataset? Also are your image bird eye or satellite view? $\endgroup$
    – Cloud Cho
    Mar 29, 2022 at 16:39
  • $\begingroup$ My augmented has 10 k, original has 1k. My images are from monocular camera on a factory press $\endgroup$
    – devman3211
    Mar 29, 2022 at 19:02
  • $\begingroup$ Thanks for the information. How many different class in you 1,000 original image? If your image collection at the factory pressing machine, is the photos of metal surface? $\endgroup$
    – Cloud Cho
    Mar 29, 2022 at 21:33
  • $\begingroup$ They are metal and plastic surfaces. So to clarify I have 25,000 images augmented total and 3,000k images original. There are 3 classes. $\endgroup$
    – devman3211
    Mar 29, 2022 at 21:41

2 Answers 2

0
$\begingroup$

I think it depending on your input image for testing. If your testing image is similar to original 1,000 images, which is no rotation on the view, I don't think augmented images help increasing classification accuracy.

I don't know what metal and plastic surface condition you want to detect, but 1,000 images per class would be enough for to reach around 80% detection accuracy for simple surface feature, like nick, inclusion, and etc.

$\endgroup$
0
$\begingroup$

This is a good question. The naïve view would be that there is no difference besides "there just being more images". However, this is not necessarily the case.

The answer depends on how "true" your augmented dataset is to your original dataset. In some cases, the augmentations are less representative of the original images. In that situation, not including the original dataset in your training may be harming your training/model performance more than "there just being more images". You would also influencing the quality of your training.

In some cases, you can tell a priori if the augmentation is representative of the original. For example, a a car vs. truck dataset that is flipped horizontally is probably representative of the original.

In situations where you cannot tell a priori, you can train a model with the original dataset and another model with an equal number of augmented images. In situations where the augmentation is "true", there should not be a huge difference in model performance. (With small datasets, the fluctuations in performance can be "huge", and this would not apply).

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .