Influence of noise polimago

polimago
#1

So this here is a bit more of a theoretical/fundamental question; what is the influence (if any) of noise on polimago training and the quality of the subsequent training?

Specifically, I want to increase the camera gain for a low light application, but since it can be a lot of work to thoroughly test the possible differences between the two situations (dark not noisy image vs brighter noisy image), perhaps some general predictions can be made?

for instance, I can imagine that the noise from the higher gain is more of an issue at smaller training sets. Is this the case?

#2

It depends on the parameterization to some extend. If you make the classifier more sensitive to fine details, for instance with many “a” in your pre-processing code, then yes, noise can be an issue. If you want to detach the classifier from noise, use one or two “p” in your pre-processing code to smooth the image a bit.

In general, I’ve not seen noise to be a noteworthy issue with Polimago, judging solely based on my personal experience. That said, the images should probably not be of abysmal quality, in regard to noise. :slight_smile:

#3

Well, the gain isn’t that high, but there is an increase in noise. But the tradeoff was motion blur, noise or dark image. So I was kind of trying to find out which would hurt the classifier the most. Thanks! :slight_smile:

#4

Motion blur might create a bigger issue than the noise I think. I’d trade some noise for less motion blur. With noise, the overall appearance of a given object might only be changed slightly, but motion blur can change it more drastically and is less deterministic (not to say that noise is deterministic :smile: ).

1 Like
#5

My gut feeling was the same, so it’s good to have it verified.

Also, I compensated with a significant increase in sample set from (280 per category)