Monday, November 17, 2008

DxO Reveals Its Image Sensor Metrics

Yahoo: At Intertech-Pira Image Sensor 2008 conference in San Diego, DxO Labs today unveils www.dxomark.com, a new website claiming to present objective metrics of sensor performance for a variety of cameras measured directly on the RAW image. “There are many valuable resources reviewing the image quality of digital cameras, but none of them consider the actual RAW signal straight from the camera sensor,” explains Nicolas Touchard, VP Marketing, DxO Labs Image Quality Evaluation business.

At first I liked the idea, but after a brief look on the web side, the proposed metrics look to me partially wrong and partially unnecessary complicated.

I can understand the rationale behind Nokia SNR10 figure to measure sensor's low-light performance. It represents the minimum illumination to achieve SNR10 after color correction matrix (CCM) bringing sensor's colors to the standard set. Being incomplete as it is, SNR10 figure is simple to use and compare. Talking about DSLR sensors, I would complement it by SNR100 or SNR200, for example, showing the minimum illumination to get near perfect picture with standard colors. Such a simple figure is full of omissions, but it's very synthetic it terms it combines a lot of sensor parameters in a single, simple to use number.

Back to DxO approach, there is no such a good synthetic number as SNR10. It's more complex as there are few different numbers making it harder to compare. For example, what if "color depth" figure is excellent, but colors are wrong and suffer from metamerism? Talking about DR and ISO measurements, how do they compare visibility of different noises? One sensor might be dominated by nicely scattered Gaussian noise, other might have mostly row or column noise, yet another one might suffer from "salt and pepper" noise. All these noises have very different visibility.

Not that Nokia SNR10 has less problems, but at least it's much simpler and much more synthetic. So if I would propose DSLR sensor metric, I would go for SNR100 measured according to Nokia procedures.

What I liked though is the database with quite a detailed measured data. The pure measured data is really valuable. Kudos to DxO for making this openly available!

No comments:

Post a Comment

All comments are moderated to avoid spam.