List - Image/ Video corpus for compression

[khavish]

While working on compression , one of the issues is to find good modern test images or test videos.As a result i think it would be a good idea to compile a list.

Info on Internet use vs Camera Images

Color images(Internet use):

Jyrki Alakuijala image corpus

WyohKnott image corpus

Camera images

Rawzor test images

Raw Pixels

Medical images

Silesia compression corpus

Video

Xiph.org Video Test Media [derf's collection]

EBU Test Sequence ( look for "Public Test Sequences" on the site. You have to register to get FTP access.)

Elephant dream lossless video and audio

Kindly suggest some more in the comments.

[LucaBiondi]

I agree i think we need a common corpus too
Not only for images
Luca

[Jyrki Alakuijala]

Kindly suggest some more in the comments.

I like to think that there are two main kinds of photographic corpora from the image compression viewpoint. First kind are the corpora that are subsampled and possibly enhanced, i.e., modeling professional internet use. Second kind are photo corpora that are camera resolution, modeling immediate storage and archival use. The two corpora you posted have subsampling and are more useful for modeling the internet use. This tends to make them relatively sharp, i.e., more high frequency content (more difficult to compress), but low on noise and interpolation/reconstruction artefacts (less difficult to compress).

The following two are two relatively good camera resolution corpora:

http://imagecompression.info/test_images/

ICIP corpus https://jpeg.org/items/20151126_icip_challenge.html

Previously on encode.su it has been shown that some of the images in ICIP have been pre-compressed in JPEG. Those images should be left out from further codec analysis, and especially so at high quality/bit rates or lossless compression.

The images with camera resolution tend to have artefacts of the reconstruction algorithm in the final reconstructed image. These artefacts have correlations. In the lossless world, my experience was, that it is possible to get another 10 % lossless density improvement by targeting such correlations, but the gain of such approaches disappears if a camera resolution image is subsampled. I'm not personally convinced that it is a good idea to try to model reconstruction artefacts of a particular algorithm in a data compression algorithm. I like to think that reconstruction algorithms are adjusted from one camera generation to the next, while an image compression system should ideally have a longer life span.

[boxerab]

For video, the European Broadcast Union (EBU) provides a public test set for free:

Go to https://tech.ebu.ch/testsequences

and look for "Public Test Sequences" on the site. You have to register to get FTP access.

[khavish]

I added some more public corpus for videos

[boxerab]

Great. (There is a typo: EDU should be EBU )

[khavish]

Great. (There is a typo: EDU should be EBU )

Typo fixed.....it would be great to have more sources for medical and camera images ...

[boxerab]

For medical images, there are quite a few here: http://www.dclunie.com/
in the "images" section, but many of these may be in DICOM format, so you would need
a DICOM toolkit to extract the images.

http://www.dcm4che.org/ (java)
http://gdcm.sourceforge.net/wiki/index.php/Main_Page (C++)

[boxerab]

Thanks for the Elephant Dream link, I've been looking for something like this for quite a while.

[thorfdbg]

/* duplicate */

[thorfdbg]

For video, the European Broadcast Union (EBU) provides a public test set for free:

Go to https://tech.ebu.ch/testsequences

and look for "Public Test Sequences" on the site. You have to register to get FTP access.

The EBU test sequences are quite ok, though you have to be a bit careful as some of them have been generated from 422 or 420 subsampled material. This is the right thing to do if you're compressing video and transmit subsampled - quite typical for broadcasting (and the job of EBU, of course), but not ideal for imaging.

I don't think ICIP images have been pre-compressed, at least not that I'm aware of. At least, I have the "raw" originals, though I did not process them to their 444 RGB representation - or at least not all of them.

There is currently another test corpus we're playing with for JPEG XS which also includes some "freaky" artificial images, and some natural ones. Images can be shared IIRC, so we should probably put them on jpeg.org at some point.

There are also a couple of well-known MPEG test sequences like "crowd run" or "park joy", though they are managed by WG11, not WG1 (and particularly, not myself). "crowd run" is quite often used, though I personally would probably argue against it - it is pretty noisy, film grain all over, but 10bpp. Digitalized analog film. This was ok as a workflow ten years ago, but today it's all digital recording to begin with, so the quality you get is typically better than that.

Oh well, that's a WG11 issue probably...

[Jyrki Alakuijala]

I don't think ICIP images have been pre-compressed, at least not that I'm aware of.

https://encode.su/threads/2392-JPEG-...made-available suggest that 2 of images are very likely precompressed as JPEG, and 2 others might be.

It was also discussed that the acquisition technology (digital cameras) of ICIP 2016 are 10-15 year old models.

Why does the age of the acquisition technology matter? Because at camera resolution some of the correlations come from the image reconstruction algorithm -- the image is not a one-to-one mapping from the ccd recordings, but has to be computed. This computation generates synthetic correlations in the final reconstructed image. Newer generations of devices may have improvements in these algorithms and produce different kinds of correlations. I think this effect is easiest to observe with lossless compression. Some algorithms (I think BCIF mostly, possibly also some with complex context modeling) can model the particularities of some reconstruction artefacts, and such algorithm get an additional advantage at camera resolution. After resizing they can lose their advantage.

Even considering these two beauty mistakes in ICIP 2016 image set, it still creates an interesting and important benchmark for image compression.

[mpais]

Repository with a large collection of raw photo samples from various camera manufacturers:
https://raw.pixls.us/

Usually between 8 to 16bps, various CFA filter layouts.

[Kennon Conrad]

The kodak image set is here: http://r0k.us/graphics/kodak/. It's not so popular these days.

[boxerab]

Elephant Dream torrent doesn't work (post is 11 years old). So, should probably remove this.

[boxerab]

For DNG images, the excellent Graphics Magick program can read and convert to other formats.

http://www.graphicsmagick.org/

[thorfdbg]

Why does the age of the acquisition technology matter?

It depends on the use case you have for your algorithm in mind. If this is for digital photography compression, then preserving the film grain does not matter, and an attempt to do so might be a burden to an algorithm that is otherwise suitable for the problem at hand. IOWs, if you are looking for an algorithm for image archives, such images may matter. If you don't, then it does not matter. The question is always which conclusions you draw from the result given the images you have. Just plotting PSNR curves and then taking the average may tell the wrong story if you do not select the images according to the use case.

[khavish]

Elephant Dream torrent doesn't work (post is 11 years old). So, should probably remove this.

You can find the direct download links for elephants dream and other lossless video source on https://media.xiph.org/

[encode]

I have my own GARTEST:
https://encode.su/threads/1659-BIM-(...ll=1#post34380

[pothos2]

Pictures:

https://web.archive.org/web/20140625...om/bitmap.html

http://imagecompression.info/gralic/LPCB.html

[Alexander Rhatushnyak]

Previously on encode.su it has been shown that some of the images in ICIP have been pre-compressed in JPEG.

Here: https://encode.su/threads/2392