Is it effective to include an affine conversion of a single image in the data that Azure's Custom Vision will teach you?

Asked 2 years ago, Updated 2 years ago, 50 views

Now I'm starting to use Azure's Custom Vision.

I learned about 30 images, but the accuracy was not high yet, so I'm thinking of learning about 1000 images to improve accuracy.

At that time, I thought that I would affine an image and inflate several images.
However, I was wondering if such geometric image data has already been padded in the Custom Vision algorithm.

If anyone understands, please let me know.

azure

2022-09-30 13:59

2 Answers

The algorithms used internally are unknown, but affine transformations may improve apparent accuracy, but I can't say whether they will actually work.Originally, even a small amount of data seems to be efficient, so it may be done internally...

Why don't you follow these guidelines first?(It seems that 50 per label is the starting point.)


2022-09-30 13:59

Dear kosmos.ebi, Thank you very much for your reply.Thank you from another account for the trouble with your account.

Affine transformations may improve apparent accuracy, but I can't say whether they will actually work.

In fact, 30 learning images were padded to 300 (geometric conversion plus random contrast and saturation adjustments), and after additional learning, the accuracy increased noticeably, and only large numbers (such as 99.99 percent) were returned.

Originally, even a small amount of data seems to be efficient, so it may be done internally...

You're right.Even with about 10 pages of learning, it is possible to infer that data augmentation or equivalent processing is being carried out internally.

For the time being, I think I will proceed as it is because the function has improved just by looking at the results.
Thank you very much for your reply.


2022-09-30 13:59

If you have any answers or tips


© 2024 OneMinuteCode. All rights reserved.