I went to Monument Valley last weekend. It was just a quick trip, and with non-photographer friends along, I mostly just played tourist. I did take one camera, one lens, and no tripod. The camera was a 4000 by 6000 pixel 35 mm format digital. I amused myself by snapping off 6 to 12 picture panoramas.
When I got home, I fired up a stitching software package called Autopano Pro 2. I’d never used it before, and I was truly amazed at what it could do.
The first revelation was its ability to accurately automatically detect groups of photographs that needed stitching together. It doesn’t do this by analyzing the images; instead it looks at when the photographs were taken, the exposure information, and the focal length of the lens. When I took the pictures I didn’t know how the program worked, and it still did a pretty good job; now that I understand it I can easily do things (like changing the focal length of a zoom lens slightly between panoramas) that should make it essentially perfect.
The second surprise was the quality of the results. When I had done panoramas before I had always used a tripod, and usually I went to the trouble of adjusting the camera so that the pivot point was at the nodal point of the lens. Even so, the stitching software left artifacts that took a lot of manual cleaning up. For distant scenes, with the camera handheld, Autopano did a flawless job. There were some problems if the foreground was too close to the camera, but it’s hard to blame that on the stitching software, since I didn’t use a tripod.
The combination of ease of use and quality of the results has given me a new perspective on stitching. I don’t think it’s just for panoramas any more. Consider the numbers: holding the camera that I used vertically, making three exposures, and assembling the result into a 35 mm shaped horizontal gives a 6000×9000 pixel image. If you’re hand holding it, you probably going to have to crop a little, so you’ll have maybe 5500×8250 pixels. Holding a camera the same way and doing two rows of three images each gives you (figuring in some overlap) a 9000×13500 pixel image. Both of those resolutions are seriously into scanning back territory.
Let’s take a moment to review scanning backs. They’ve been around since the early 1990s, a time when rectangular image sensors captured less than two megapixels. The idea was that, although it was incredibly expensive to make a high-resolution sensor that could capture an entire image at once, it wasn’t a big deal to build a line sensor with thousands of pixels. If you had a subject that wasn’t moving, you could use a motor to slowly move the sensor across the entire image. Several companies built scanning backs that slid into 4×5 cameras like film holders. It was a little like having the guts of a flatbed scanner in your camera. There was an umbilical that plugged into a box that you put under the tripod, and the box had to be connected to a computer that you took into the field. Capture times were measured in single-digit minutes. The results were spectacular. Quality exceeded what could be obtained with 4×5 film by quite a bit. The long exposure times limited the choice of subject matter, and the requirement to tote a computer along made field work slow and awkward.
Today, the standard resolution for a scanning back is 6000×8000 pixels. If you need higher resolution and have $23,000 lying around, you can get a 10200×13600 pixel back.
There are two big differences between scanning and stitching that make comparisons more complicated than just counting pixels.
The first favors scanning: each pixel in a scanned exposure is a combination of independent red, green, and blue sensing elements, as opposed to the pixels in an instant capture that are interpolated from the Bayer pattern in the sensor. I discussed that issue several years ago, and I figured that, to get the equivalent of real three-color pixel capture, you should divide the number of pixels in an instantly captured image by two, the equivalent of dividing each dimension of the image by 1.4.
The second favors stitching: view camera lenses are not built to give the kind of resolution demanded by high performance scanning backs. It’s axiomatic in lens design that as coverage increases, resolving power (measured in line pairs per millimeter) decreases. A bigger capture area and a bigger lens will get you more good pixels, but the increase will be less than proportional to size. When you use stitching to get high resolution, you can use a small lens with high resolving power; you get the increased resolution by using the lens over and over for each exposure, rather than the much more challenging procedure of trying to get a lens that can resolve the entire shot at once.
For studio use with no people in the shot and continuous lighting (no strobes), the scanning back still has a place. But in the field, where the bulk of all the equipment you have to carry with you (camera, back, tripod, cables, electronics box, computer, etc.) can really limit your mobility, increase your set-up time, and send you to the chiropractor, stitching together small images is getting increasingly attractive. As I found out in Monument Valley, sometimes you don’t even need a tripod.
In the past, my reaction to the difficulties associated with using a scanning back was to just forget about really high-res images. With the twin improvements in instant-capture sensor resolution and stitching software, I think I’ll change my mind.
However, I do have a problem with some of my Monument Valley pictures: wall space. There’s an image I like especially well. It’s composed of twelve verticals arranged horizontally and it’s 15000 by 6000 pixels. At 360 pixels per inch, that’s 17×42 inches. It’s a nice picture, but it’s not good enough to turn that much wall over to it, and if I print it smaller, you won’t see all the detail.