Return to Ray Tracing Assignment Overview
 

Depth of field is implemented by giving the "eyepoint" an area, much like soft shadows work by giving point lights an area.  Here, the size of the aperture (for our purposes, the lens size) is calculated by the f-stop--just like in a real camera.  Larger apertures are denoted by smaller f-stops.  On a normal 35mm camera, the largest aperture is usually around f-2.8 or f-3.5, where the smallest aperture is usually f-22 or even f-27.  The size of the aperture determines the depth of field for a scene.  That is, a large aperture will result in a small depth of field...or a smaller z-distance that is in focus.  When using a real camera with a large aperture, it's very important to make sure that you focus on exactly what you want to have in focus, because everything else will be slightly blurry.

Our implementation takes the normal eyepoint and shoots a (fake) ray to the focal plane for a particular pixel (fake. adj: it doesn't return a color).  That intersection is noted.  The real ray (the one that returns a color) has origin not at the eyepoint, but some random point on the lens.  The ray is shot to the intersection point on the focal plane that was noted earlier.  This means that everything on the plane of focus will be hit where it originally would've been hit with a pinhole camera model. The further away from the plane of focus an object is, the more "incorrect" point it will be intersected--resulting in the blurring effect we're after.

My biggest issue with this implementation was that I wasn't paying enough attention to the physics of real cameras; I thought that I wasn't getting the correct results, when actually I was.  In my raytracer, the lens radius is defined in terms of meters, meaning that for the purpose of depth of field, my scene is also described in terms of meters.  My Cornell box scene of previous assignments is 10m cubed.  I have to stand 20m away to fit the entire room in the viewplane of the camera (this seems realistic).  On most 50mm camera lenses (which is what we're modeling) there is no option to focus at 20m...20m means infinity.  You just can't get any depth of field effects at that distance with realistic apertures.

To illustrate depth of field in my raytracer, I brought the objects closer to the camera.  The somewhat lame scene below (sorry, I don't have time to model something cool) have objects at 0.25m, 0.5m, 1m, and 50m away from the camera lens (labeled).  This first set of four images all have aperture sizes calculated for a 50mm lens at f-3.5 (the largest aperture on my personal 50mm lens).  Each was set to focus on a different z-plane.  All the spheres have radius 50mm except the green one which has radius 12 meters.
  

This second set of images are all focused at 0.5m (the yellow sphere) but have varying f-stops.  These are normal f-stop values found on typical 50mm lenses.  Notice how at f-22 almost everything in the scene is in sharp focus.  This is do to the fact that at f-22 the diameter of the aperture is just a little over 2mm...almost a pinhole.  Compare this to almost 18mm, the size of a 50mm lens aperture at f-2.8.
 

Ah, back to my Cornell box problem...  The great thing about computers is that they don't have to obey the physical restrictions of real cameras.  The image below is focused at 22.75m (the location of the glass sphere) and set to f-0.025...meaning the diameter of the aperture is 2 meters wide!  It's too bad there isn't a real camera that big.  (Actually, in real life, this problem is solved by zoom lenses...like the large ones you see professional photographers have at sporting events.  These can have a focal length of 1600mm and usually just one f-stop, something like f-16.  Meaning that effectively the aperture diameter is something like 100mm.  Zoom lenses--by their nature--actually shorten the depth of field of an image more so than standard lenses...so the blurring the would occur is actually more than what you'd expect with a 0.1m aperture.)
 

Unrealistic Depth of Field Possible without Camera Restrictions
400 x 400 JPG converted from 400 x 400 PPM output
Aperture: f-0.025 (!)
256 sample per pixel
22.923 minutes
 

As a side note, the reason the wall intersections still have this nice (ugly) square shape (instead of being blurred out of their square shape) is because my aperture is based on a square, not a disk.

One more thing:  Ever since project 2 I've had this "vertical-field-of-view" parameter in my raytracer.  I never knew what this actually should be set to for realistic cameras.  For this project I looked it up.  The vertical-field-of-view for a 50mm lens with film size 36mm x 24mm is 29.9915 degrees.  So, that's the default setting now.  Like I mentioned above, it seems realistic, because in order to fit the 10m tall Cornell box into my scene the lens has to be 20m away.

Executable compiled with Microsoft Visual Studio .NET Professional and run on a Dell desktop with an Intel Pentium 4 1.8 GHz processor with 1.0GB of RDRAM running Microsoft Windows XP.