Content aware image resizing from Siggraph 07

A common bugbear encountered when working with diverse collections and images is the inability to gracefully created resized versions. We have never found a suitable solution to creating thumbnails of our collection for the OPAC and Design Hub – the current solution is to take the existing large image, resize it to be 500 pixels on the longest side, then take a square from the middle 400 pixels, and resize the square to 80×80 pixels, including any white space borders. This is run as a batch process. Whilst this works for most rectangular images it still has the unintended side effect of lopping off heads and feet, and on rare irregular shapes such as very long artworks, the thumbnail is virtually useless even for quick object recognition tasks.

Here are some examples –

But here, in a presentation from Siggraph 07, is a fascinating potential solution. It is quite amazing and by reducing or expanding an image based on ‘content’ has very interesting implications for intellectual property legislation. In many ways it does what MP3 compression does (poorly) for audio, intelligently remove the bits of the image that are least recognised by the viewer. In so doing it makes assumptions about the overall image – and how we ‘see’ images.

4 replies on “Content aware image resizing from Siggraph 07”

Wow, that’s really, really clever. I wonder what the overhead for a machine is to scale those dynamically, or if the least-important lines could be somewhat pre-computed and stored as meta data (including the “hints”)? Then a browser that supported retargeting could do its thing, and everything else would ignore it…

If you are looking for a software to try out seam carving, take a look at

The program that you can download there (for Windows and Linux, and free) is already highly optimized concerning speed, and apart from enlarging or decreasing image size you can also use masks to protect or delete certain parts of your image.

Have fun,

Comments are closed.