I listened to a lot of Bowie during this project. It was roughly period appropriate.
Input: The classic Aladdin Sane cover:
Hockney-Style Composite Output:
David Hockney is an artist who has dabbled in lots of different media and disciplines, from painting to stage design to, briefly and famously, photography.
Though Wikipedia says he eventually grew frustrated with the camera’s single-eye approach to a scene, his efforts to compensate for it through collages and polaroid composites are well documented.
While Hockney’s compositions are created from many different images with all the quirks and subtleties of a human hand holding a 35mm or polaroid camera, my idea was to recreate the style of his collages and composites using a single source image.
Hockney’s collages are composed of lots of photos of the same subject or same scene, layered and mixed and jumbled to give textural comment. The polaroid composites are orderly layouts of polaroids taken of the same scene at different times and slightly different points.
To approach this, I built a custom library centered around the abstraction of ImageTiles. These are sub-images of the source file that have lots of different properties: show, don’t show, zoom in, zoom out, bright, darken, emphasize, scatter, and room for lots more.
After a semester spent mired mostly in C++ with a little Python, I took the liberty of doing this in Java, which was… really nice, actually. It was fun to be able to get ideas out without tripping over syntax. I will get around to learning Python properly someday, but that is not this day! The source code is here: https://github.com/ecurtin/PhotoCollage/
Using this custom library, which relied on some oldie but goodie and MUCH MORE NICELY FORMATTED code from my dear old writes-Java-like-its-70’s-Fortran computer graphics professor, I built an application to make collages and an application to make composites.
The applications follow roughly the same format:
- Input the image
- Split it into ImageTiles
- Pipe these imageTiles through a number of transformations including
- Scatter: move the coordinates for where the tile will be rendered to a random location a random distance away
- Show/Hide: For the collage, randomly show or hide the tile based on a density parameter
- Brighten/Darken: make the image randomly a little brighter or a little darker (I found a little goes a looong way for this one)
- Zoom: randomly zoom the tile in or out.
- Others! It’s super easy to write a method to do something to a single tile and simply add it in the main run loop of the program.
Bonus, although I haven’t gotten around to it, this is embarrassingly parallel. I could speed this up IMMENSELY by generating a bunch of worker threads to each work on their own tile or small set of tiles.
And speaking of things I haven’t gotten around to yet, there’s lot of room for improvement here, especially in the collage. A few ideas:
To increase the analog or human jitter realism:
- Different kinds of exposure modification beyond just lighten/darken
- Random rotation within a specified range
- Color distortion to simulate differences apparent in the chemical development process
To further the aesthetic qualities:
- Apply scatter based on heuristics instead of pure probability. Move in the direction of the gradient of the tile?
- Apply heuristics and/or computer vision techniques to detect interesting features and large swaths of similar space, emphasizing and de-emphasizing as appropriate (is de-emphasizing a word?)
To make the application more friendly/usable, since apparently nobody likes one-liner shell scripts with ~18 anonymous numerical parameters
- An actual GUI! Something I haven’t had to think about since I started doing Systems! Maybe github.io? Maybe just a .jar?
- Paralleliize (how on earth do you spell that?) for speed
- Perhaps an interface where the user can graphically select regions of interest as input
Finally, Professor Irfan Essa: