I think some posters (BartvanderWolf, pegelli, b_z and image66 among others) haven't understood Marc Dubovoy's excellent article because they are confused about what microlenses accomplish.
Microlenses improve the angular response of a sensel, but only up to a limit.
The marginal rays of the light cone projected by a fast — e.g. f/1.2 — lens are quite tilted.
A pont source of light, e.g. a small LED photographed from a fairly large distance, stars in the night sky etc. form, when in focus, a point on the imaging sensor or film.
When out of focus, the intersection between the lens' light cone and the imaging plane forms a disc, not a point.
Depending on the degree of defocus, that disc can obviously have various diameters.
If the degree of defocus is small, the disc diameter might be smaller than the diameter of the circle of confusion which figures in depth of field calculations.
Regardless of the OOF disc's diameter, the angle of the marginal rays relative to the sensor doesn't change. This obviously implies, as Mark Dubovoy points out, that the diameter of the circle of confusion recorded by an imaging sensor is also affected by the sensel's acceptance angle, and that depth of field, in turn, must be affected by the sensel's angular response.
How can we estimate that acceptance angle with any degree of reliability ?If, despite the microlens' best efforts, the marginal rays are too tilted to reach the photodiode at the bottom of thick sensel structure, these light rays can be considered to be non-existent for imaging purposes.
If these marginal rays cannot reach the photodiodes, the recorded diameter of the disc formed by an out of focus point light source will be reduced.
A simple way to assess the critical incidence angle — that is, the acceptance angle — above which the rays cannot be recorded anymore is thus to measure the diameter of the out of focus disc formed by a point source.
Film doesn't have any problems recording even very tilted light rays. Due to the very geometry of the lens' light cone, the diameter of the OOF disc must be directly proportional to the lens' f-stop a.k.a. aperture setting. On film, doubling the aperture thus necessarily doubles the recorded OOF disc diameter.
With some digital sensor designs, acceptance angles can be quite limited. At some point, the linear relationship between the lens aperture and the OOF disc diameter recorded on the picture must then break down, as the tilted marginal rays cannot reach the photodiodes anymore.
A quick measurement of the dimensions of the OOF discs visible in the
test picture of the Canon EF50mm F/1.2L lens on
Photozone.de show that with the Canon 5D Mark II's CMOS sensor used in the test, the diameter of the OOF disc increases, as expected, when the lens is opened up from f/4 to f/2, but then plateaus at a diameter geometrically corresponding to about f/1.5, even when the lens is opened up to f/1.2
This indicates that the acceptance angle of the Canon 5D Mark II's sensor is limited to about arctan(1/(1.5*2)) i.e. about 18.4 degrees from the perpendicular, and that a f/1.2 lens thus performs on the 5D2 essentially as a f/1.5 lens as far as actual lens speed, DoF and bokeh are concerned.
To fully record the bokeh and "draw" of fast lenses like the Canon EF50mmF1.0L, EF50mmF1.2L and EF85mmF1.2L, ir thus seems that one will have to use a film-based Canon EOS body, instead of a Canon DSLR.
An interesting pint of comparison to bring up here might be the — presumably identical — sensels used in the CCD sensors equipping the Leica M8 and M9.
CCDs don't need the multiple transistors surrounding each photodiode of a CMOS sensor architecture. As such, CCDs don't require the multiple metal layers of a CMOS needed for the transistor's signal lines. The resulting CCD pixel stack is typically much more shallow than a CMOS sensor's tunnel-like architecture.
With a CCD, the distance between the microlens and the photodiode can thus be made closer than with a CMOS sensor, and the microlens' acceptance angles can thus be much wider.
The marginal ray of the light cone of a Leica Noctilux f/1.0 lens is about arctan(1/2) = 26.6 degrees. According to its
datasheet, the Kodak KAF-10500 CCD used in the Leica M8 still has an angular response, at an angle of 27 degrees, of about 70% of the peak response.
Unlike the DSLRs, the CCD-based digital Leicas thus seem entirely able to record the light rays of the wide light cone emanating from Leica's super-fast Noctiluxes.
As for b_z's assertions about acceptance angles, using his NEX-5 and a 40mm F/1.4 lens as examples, they are flawed for several reasons:
- b_z doesn't properly consider the distance between the sensor and the lens' exit pupil, from which the light rays can be considered to emanate. That exit pupil distance is unknown. If it is 40mm, the chief ray's tilt angle, even in the extreme corner of an APS-C sensor, would be arctan(14mm/40mm) i.e. about 19.2 degrees, that is, quite close to the Canon 5D2's f/1.5-equivalent 18.5 degrees. As such, given the incertitude about the exit pupil distance of the 40mm lens used by b_z, his test doesn't bring any useful piece of information about the angular response of f/1.2 lenses.
- The minimal exit pupil distance relative to the image plane of most Leica-mount rangefinder lenses, including ultra wides, is about 28mm. I know of no actual instances of Leica-mount lenses with a 15mm exit pupil distance hypothesized by b_z.
- The Sony-brand APS sensor used in Fuji's X100 fixed-lens digital rangefinder camera uses offset microlenses. There is no publicly available information indicating whether Sony uses offset microlenses for the NEX series of cameras. This means that it is impossible for us to state whether the NEX-5 uses, or does not use offset microlenses. If the NEX-5 uses offset microlenses — a plausible scenario, given the very short flangeback of the NEX mount and the compactness of the NEX 16mm pancake lens — the acceptance angle of the corner sensels could be even larger than the back-of-the envelope calculations done above, rendering b_z's NEX-based analysis even more meaningless.
Let us now consider image66's assertion that the vignetting is not caused by the sensor, but by the mirror box.
If the mirror box were to cause vignetting at the center of the image where DxO performed their mesurements, then, the OOF disc of an f/1.2 or f/1.4 lens at the
center of the image would not appear as a perfect circle, but as a circle visibly truncated by the straight edges of the mirror box.
Given that anybody with an APS or full-size DSLR and a f/1.2 or f/1.4 lens can verify that, even with the lens wide open, an OOF point lioght source appears at the center of the image as a perfect circle and not as a truncated one, image66's assertion can be dismissed out of hand.
As a camera geek, I applaud Mark Dubovoy for his very insightful article, and look forward to DxO publising their detailed findings as to how they assessed the amplitude of the ISO correction camera manufacturers automatically apply to compensate for the non-recording of the marginal light rays emitted by fast lenses.
