Where do you recommend getting "standard" images for testing algorithms?

Question

I've noticed that there are a LOT of websites which claim to have a standardized repository of images (for image processing algorithms). However, I don't know which ones are actually trusted by the research community, and which are not. I know some file formats hint of untrustworthiness (e.g. never use a JPEG since its compression is lossy, and it is probably not the original image). Many of these repositories appear to have the same image, but their actual files do not match pixel-for-pixel.

So, what is your recommended repository for getting standard images for testing image-processing algorithms (e.g. images like cameraman, boat, airfield, peppers, pirate, etc.)

Answer 1

Research communities don't "trust"; they take a set of data and consider its properties (sample size, representativity, quality, systematic biases...) and then pick that. There's no "universally trusted" thing, as a dataset that's perfect for one purpose might be very wrong for another.

So, you'll have to work this from the other side up: pick a set of images that well represent what you want to do! Is it improving the quality of compressed photographs of cattle? Different dataset then you'd use to assess a denoising algorithm for architectural photography!

If you want to do that works on most photographic content, for example, you'd probably end up using the same dataset as the JPEG used to develop their compression standards (of the same name); not only has that been proven to actually be a useful set, but also it allows a fair comparison to established methods that use the same.