Making Sense Of Sensors; Interpolation Before And After The Fact

Ever since Foveon, Inc., based in Santa Clara, California, announced their unique new image sensor design back in 2001, it has been the subject of some controversy. Foveon's initial promotional campaign proclaimed the virtues of their invention in glowing terms while denigrating the competition, with the predictable result being a background level of skepticism that persists to this day. Over the past 5-6 years, Foveon has quietly continued to develop their three-layer sensor, and teamed up with Sigma to create a series of increasingly sophisticated D-SLRs employing their latest sensor iterations. The culmination of this effort is found in the current Sigma SD14, which incorporates the Foveon X3 sensor, a three-layer, 21x14mm CMOS with 4.7 million pixels per layer, giving a claimed total of 14.1 megapixels and a lens focal length multiplication factor of 1.7.

Both Foveon and Sigma staunchly defend their right to spec the X3 sensor at 14.1 megapixels. International standards groups like the ISO have yet to issue specific guidelines on presenting pixel data, but the CIPA (Camera & Imaging Products Association) guidelines allow all the pixels in a three-chip camera to be counted, and this should logically apply to a chip with three sensor layers as well. However, the real question is not how many pixels the Foveon X3 sensor can rightly claim, but how well it performs, in practical terms, compared to its leading Bayer-pattern rivals. To find out, we made tests using the Sigma SD14 and two of the most highly acclaimed, best-performing prosumer D-SLRs in the business--the 10-megapixel Nikon D200 and the 12.8-megapixel Canon EOS 5D.

Our Tests: Trying For A Level Playing Field
Devising a comparison test protocol in the digital age is more challenging than one might imagine. To begin with, the three test cameras have different formats--the Canon EOS 5D has a virtually full-35mm-frame (35.8x23.9mm) CMOS sensor while the Nikon's DX CCD sensor measures 23.6x15.8mm, giving a lens multiplication (cropping) factor of 1.5. To capture nearly identical images with all three cameras, we therefore had to shoot the test pictures at different distances, varying them by the ratio of 1 to 1.5 to 1.7 for the Canon, Nikon, and Sigma respectively. Minimizing the lens quality variable was easier. Sigma offers a 70mm f/2.8 Macro in all three mounts, and it's one of the highest quality lenses in their line, so we were able to shoot all the test pictures with the same lens, albeit in different mounts. The last problem proved to be the stickiest, namely how to get comparable files for making straight, unmanipulated prints from all three cameras? To achieve this, we shot all three cameras in raw at maximum resolution, converted the files to TIFFs using each camera maker's proprietary software, and printed them out directly without any manipulation, sharpening, or color correction.

All prints were made with an Epson 3800 inkjet printer using Premium Luster paper and Epson 3800 profiles at 1400dpi. The final TIFF files from all cameras were approximately 13x19" at 240dpi, and all were converted to the Adobe RGB (1998) work space prior to printing them for evaluation. Because the original Sigma TIFF files are smaller (2640x1760, or 13.3MB at 8 bit), corresponding to a 7.3x11" print size at 240dpi, they had to be interpolated, or resampled up to make the test print. Likewise for the Nikon NEF files, which at 3872x2592 pixels (a 28.7MB file) yielded a 10.8x16.1" image prior to resampling. The Canon image (4368x2912, or a 36.4MB file) came closest to the target print without resampling at 12x18" at 240 ppi. All resampling was done using Photoshop CS3 with bicubic upscaling. We believe we've come about as close as possible to minimizing the variables, creating a reasonably level playing field, and giving all contenders a fair shot.

The Proof Is In The Prints
Canon EOS 5D
Nikon D200
An outstanding result, with pleasing, well differentiated skin tones and colors in background flowers. Sharpness is extremely high--note details in model's upper lip and eyelashes, highlights on glass vase, edge detail in blouse and on fabric swatches.
Excellent overall color reproduction with more natural-looking skin tones than the other cameras. Overall contrast is lower, but detail rendition is still impressive. Minor color artifacts are visible in details of blue fabric swatch.
Sigma SD14
Slight overall greenish cast is especially noticeable in skin tones and blues, but sharpness and detail rendition are outstanding and on a par with the Canon EOS 5D--note texture of blouse material, excellent details in fabric swatches, extremely well defined skin textures and hair. Contrast is high but this enhances perceived sharpness. (Note: Greenish cast perception was enhanced by test target results, not shown.)
Conclusion: We'd have to declare the Canon EOS 5D print the winner here on the basis of overall quality. However, many would pick the Nikon D200 print as the most pleasing of the three for this kind of subject. The Sigma SD14 print exhibits outstanding sharpness and detail rendition--a tad better than the Nikon and almost equal to the Canon. But the greenish cast and high contrast make it less appealing. Needless to say, both of these defects can easily be corrected by tweaking the curves, gamma, and color balance, but that would not be in accordance with our test parameters.
Photos © 2007, Jason Schneider, All Rights Reserved

The Test Shoot
All test pictures were shot at f/8 at ISO 100 at the Auto White Balance (AWB) setting, and all the cameras were mounted on a heavy-duty Davis & Sanford Carbonlite X10 tripod. We focused manually with the aid of a magnifier, and to minimize variations caused by focusing errors, we shot several frames at slightly closer and farther distances than the viewfinder-indicated best-focus point and selected the sharpest image files for printing. There are slight variations in image size and framing because the calculated shooting distances did not correspond precisely with the angular coverage of the sensors, but we judge these to be insignificant in evaluating comparative imaging performance.

To provide some scientific background data for our test, and because we were frankly curious about the results, we also shot a series of pictures of a group of color and black and white resolution test targets with all three cameras. We thought it would be interesting to see whether the differences in black and white and color resolution between the Foveon X3 and Bayer-pattern sensors predicted by some experts would actually be visible. Note: The results of the test target series are not incorporated into our findings and do not affect the results of our practical comparison test.