Why does the technique of the previous post work? I think it may be a combination of the following:

• I didn’t use liveview to focus the D3s reference image, so it could be slightly less well focused than the Sony images. I will use liveview in future testing, although it increases noise by heating the sensor. Should I make the reference images by focusing with liveview, then turning it off, letting the sensor cool down, then making the exposure? Or is that tilting the playing field too much?  [Addendum: I made some test images, and that's not it; the autofocus pictures are as sharp as ones focused manually with the aid of liveview.]
• The antialiasing filter in the D3s blurs more (measured in microns, not pixels) than the AA filter in the Sony’s, because the pixel pitch is larger.
• The processing done by Perfect Resize artificially sharpens the image. If that’s the case, does that processing adversely affect some subject matter? It’s not just the usual sharpening, because I can’t duplicate it with the Nikon images. It will take some testing to find out.

I wonder how far I can push the ISO and still get high-quality 5 megapixel images? I wonder what would happen if I resized the NEX-7 images to 12 megapixels and compared them with full frame D3s images? More to come.

I have exciting news. At least, it’s exciting to me. Your reaction may be, “Kasson, you really need to get out more.” Anyway, after many unsuccessful attempts, detailed in the previous posts, to tame the NEX-7 high-ISO noise by reducing the resolution of the image, I finally found something that works.

Here’s the recipe. Go into Lightroom or your favorite raw converter (I’ve only used Lightroom, but the technique should work in any other raw converter) and open up the noise controls. Get rid of all, or almost all, the visable noise. Be aggressive. Let the detail suffer a bit. Then resize the image downwards. I used Perfect Resize with the default control settings.

Using this approach, I resized NEX-5 and -7 images made at ISO 3200 from their original sizes to the 2784 x 1848 DX resolution of the reference Nikon D3s, and got lower noise and better details!

The crop of the NEX-7 image looks like this:

The crop of the NEX-5 image looks like this:

The crop of the reference D3s image. very lightly touched up with Lightroom noise reduction, because it’s so clean to begin with:

In search of optimum tradeoff of noise and resolution in a 5 megapixel resize of NEX-5 and NEX-7 images, I tried a series of averaging operations before the resize. The one that worked best for the NEX-7 test image was a 2-pixel box blur in Photoshop followed by a resize in Perfect Resize using the default parameters. Even so, the results were inferior to the reference D3s image — both noisier and not as sharp.

The crop of the NEX-7 image looks like this:

The crop of the D3s image looks like this:

Included for completeness is this crop of the NEX-5 image, although the 2-pixel box blur is too broad for the 3:1 downres:

It’s a truism that sensors with larger photosites have less noise. But you can reduce noise by combining information from small photosites into fewer pixels than with a standard Bayer conversion. In my mind, this is ideally performed before demosaicing, but raw converters don’t, in my experience, allow to to choose the resolution rendered by the demosaicing operation.

With my NEX-7 studies, I’ve been wondering how much can be done after raw conversion to trade off resolution and noise. If you’ve been reading this blog a while, you know that I attacked the question by conducting an experiment. I took three cameras with APS-C sensors, and different pixel pitches. The first one, the Sony NEX-7, has 24 megapixels. The second, the Sony NEX-5, has 14 megapixels. The third, the Nikon D3s, has a bit over 5 megapixels.

“Just a minute here,” I hear you saying, “the D3s has 12 megapixels, and it has a full frame sensor.” Indeed it does, but it also has a mode that induces an in-camera crop to an APS-C-sized sensor.

To make sure the comparisons were apples to apples, I used the same lens on each camera: the AF Nikkor 35mm f/2 D. It has an aperture ring, so I was able to use it on the two Sony cameras with the Novaflex adapter.

I made three photographs, all at f/5.6 and a 50^th, with the ISO set to 3200. I focused the two Sony cameras using manual focusing with maximum magnification. I used autofocus for the Nikon, which probably slightly adversely affected the results – I probably should have used liveview. I locked the mirror up on the Nikon; the two Sony’s have no need of that. I converted the images from raw in Lightroom. 3.6, with noise reduction turned off.  Here’s the one from the Nikon:

Then I reduced the resolution of the two Sony images down to 2784 x 1848. For the images in this post, I used Photoshop bicubic sharper, which Adobe recommends for reduction. I cropped all the images to 360 pixels wide. Then I doubled the magnification using nearest-neighbor, which yields a two-by-two array of identical pixels for each pixel in the original image. The JPEG compression used for the web smooths over these clumps. Each camera had internal noise reduction turned off. Each camera was set to produce images in the Adobe 1998 RGB color space, but all images have been converted to sRGB for web presentation.

The crop of the NEX-5 image looks like this:

The crop of the NEX-7 image looks like this:

The crop of the D3s image looks like this:

There’s a lot more noise in the two Sony images, but the D3s image is softer. The bicubic sharper interpolation is sharpening the two Sony images. I used unsharp masking in Photoshop to make the Nikon image about as sharp as the Sony ones. It’s still less noisy:

I didn’t remember having trouble with purple corners with the NEX-5, so I thought I’d run a test. I took photographs of a cloudy sky with both the NEX-5 and the NEX-7. I used my worst lens for generating purple corners, the Leica Elmar f/3.8 24mm. I set it to its worst aperture for purple corners, wide open.

Here’s what I got with the NEX-7:

And with the NEX-5:

The NEX-5 is not completely free of color casts as you go away from the center of the picture, but it’s a world away from the NEX-7.

Chastened by my inability to achieve consistently accurate autofocus touchup with the Zeiss Sonnar f/1.8 on the Sony NEX-7, I made a series of 50 photographs using manual focus and analyzed the results. The only important difference between the focusing experience using autofocus override and that using manual focus on the NEX-7 with Sony lenses is that using manual focus allows you to focus using 12x magnification. I found it easier to focus the Sony/Zeiss lens with the increased magnification, but it was still difficult to get the focus point exactly where I wanted it.

As expected, the greater magnification increased my probability of success with manual focus, raising it from about a one third chance of getting a sharp image using 6x magnification to a 92% chance of getting a sharp image with 12x magnification.

Then I took the Sony/Zeiss lens off the camera, put on the Leica 24 mm f/3.8 Elmar, and made 50 more pictures. Every one was sharp. Not only that, I was able to achieve precise focus much more quickly than with the Sony/Zeiss lens, and the focusing operation was not at all stressful.

As I said in the previous post, I think the source of the less accurate, less pleasurable focusing experience on the Sony/Zeiss lens is the focusing algorithms that are built into the NEX-7. Having made so many exposures back-to-back with a lens with a normal focusing mount and the Sony/Zeiss 85 mm, I think I have some insight into what the focusing algorithm in the NEX-7 is doing. It seems that when you rotate the focusing ring rapidly, the ratio of lens movement to ring movement is higher than when you rotate the ring slowly. This may be a deliberate move on the part of the Sony engineers, who may be thinking that when you’re moving the ring slowly, you want the lens to go and some fine-tuning mode where it takes a lot of movement of the ring to get the lens to move a little bit. Maybe that would be a good thing for somebody who hasn’t spent much time using a real focusing helicoid, but it drives me nuts. I’ve spent fifty years learning how to focus really well, and I don’t want anybody to change the rules on me.

Now, let’s take a look at the flip side of NEX-7 manual focusing. With the Elmar, a wide angle lens, I got 50 successive perfectly focused pictures. I couldn’t do that with a loupe and a large-format camera. I couldn’t do that with a split-image or microprism insert in the middle of a SLR ground glass. I couldn’t come close to it with manual focusing on a modern DSLR. With the right cam in the right rangefinder camera (the odds of that aren’t good) I might be able to do it, but if the camera and the lens aren’t perfectly matched, I’d get no sharp pictures. This is spectacular focusing accuracy that NEX-7 users can rely on. It’s a major step forward.

Yesterday morning, I set out to make a series of pairs of images, letting the NEX-7 focus the first image in the pair, and manually focusing the second. The images did not turn out the way I expected. I expected to catch the NEX-7 making autofocus mistakes. Now that I understand the limitations of the NEX-7 autofocus better, I found no out-and-out mistakes at all when I used it in circumstances where it can be expected to work. I also expected that I would be more accurate manually focusing in all situations then the NEX-7 autofocus. I was wrong.

Here’s the good news. The NEX-7 autofocus is quite accurate if the entire portion of the image surrounded by the focus indication rectangle in the viewfinder is essentially in the same image plane, and possesses information of medium or high spatial frequency. If the portion of the image that the camera is focusing on contains elements of varying distances so that they can’t all be in focus, you don’t know where the camera is going to focus.

Here’s the bad news. While I continue to admire the concept of the feature of the NEX-7 that allows you to manually tweak the autofocus just by twisting the focusing ring of Sony lenses, experience has caused me to give the implementation mixed reviews.

There are two problems: the first is that the magnified image that appears when you twist the focusing ring is the 6x magnification not 12x. That makes it difficult to see when the image is in sharp focus. I looked in vain for menu settings to allow me to change this. It is possible to get the magnification to briefly go to 12x by pressing the lower button on the back of the camera while holding the shutter release half-depressed, but I couldn’t get it to stay that way for more than a second, and I took a lot of unintended photographs while trying.

When you twist the focusing ring on the NEX-7 you’re not directly controlling the focus of the lens. The connection between the ring and the focusing rack is electrical, not mechanical. When you twist the ring, you’re telling the camera firmware that you want it to change the focus distance, and the firmware then tells the lens what to do. The algorithms that do that do not provide a good simulation of mechanical focusing. A small part of this is the feel of the focusing ring, which doesn’t replicate the action a Zeiss or Leica M-mount lens. The larger problem is that the translation of ring angle to focusing is a bit strange. I don’t think it’s time lag, but I have a tendency to overshoot when making small corrections. I’ll be doing a post on this in the future.

Both the image magnification and the focus algorithm problems are probably fixable in firmware, and I hope Sony addresses them.

I made a set of 30 pairs of photographs of subjects in which the Sony autofocus should work well. In the first image of each pair, I let the camera focus. In the second, I tweaked the focus manually. In roughtly a third of the images, I did as good a job as the autofocus – maybe better in one or two. In the rest, the autofocus was better than I was.

Some examples follow.

First, here’s an image where the autofocus area, one-third in from the left and one-third down from the top, contains objects of varying distance. This photograph was made at f/2.8, handheld at a fortieth of a second. It is surprising to me that the resolution of the NEX-7 sensor is so good that the depth of field of the lens can’t span the range of distances in the autofocus area. It is also surprising to me that the shutter release is so shake-free that you can make a photograph this sharp at such a low shutter speed.

Here’s a one-pixel enlargement of the area that should be in focus. The camera, on autofocus, focused behind the tree trunk that I wanted in focus:

Here’s a one-pixel enlargement of that area when I focused. It looks the way I intended.

Second, here’s an image in which the autofocus area, one-third in from the left and one-third down from the top, contains objects which all fall within the depth of field of the lens when it’s focused on any of them:

Here’s a one-pixel enlargement of the area that should be in focus:

Here’s a one-pixel enlargement of that area when I focused. The sharpness is the same, to my eye.

I’ve read a few complaints about the NEX-7′s autofocus performance when used with the Sony/Zeiss 24 mm f/1.8 lens. I don’t think it is that the NEX-7′s autofocus is deficient compared to similar cameras; I think the problem is that the camera body and the lens are both capable of such a high level of performance that small autofocusing errors, which would go unnoticed in more forgiving cameras, are glaringly apparent.

In my experience, the contrast measurement autofocus systems on cameras like the NEX-7 don’t offer the repeatability and, in cameras built to close tolerances, the accuracy of the phase detection systems used in DSLRs. Contrast measurement autofocus systems have a theoretical advantage in that they don’t depend upon the distance from the focusing elements of the lens to the autofocus sensors reflected in the mirror being the same as the distance to the sensor after the mirror flips out of the way. However, the contrast measurement systems appear to be slower, less precise, and less accurate. Perhaps this is not a fair comparison in that my experience with phase detection systems is in quite expensive cameras, and my experience with contrast measurement systems are in cameras priced like the NEX-7 and lower.

One of the first difficulties you encounter when using autofocus with the NEX-7 is the fact that the area of the focus sensor is quite large compared to the best single lens reflex cameras. This means that it’s harder for you to get the camera to focus on what you want. The second is that the camera sometimes decides that the image is in focus when actually it is slightly defocused.

The “focus peaking” feature could in theory give you valuable feedback about whether or not the camera has focused on the object you wish it to. There are two problems with using the focus peaking feature this way. The first is that it really is not measuring whether the images in focus; it is simply measuring when the high spatial frequency content of the image exceeds a certain threshold. If your subject has no high spatial frequencies, then it’s not going to set off the focus peaking indication no matter how well it’s focused. The second problem is that, even in its most critical setting, the focus peaking feature is too easily satisfied with objects of high spatial frequency – it will say things are in focus even if they’re slightly off.

But there is a magical feature of the NEX-7 when combined with the Sony and Sony/Zeiss lenses that make all of these problems moot. You can set the camera up in autofocus mode, let it do its thing, and then, with the shutter release still partly depressed, twist the focusing ring. The image in the viewfinder is instantly magnified. You can complete focusing, finished depressing the shutter release, and you’ll have a sharp picture. This feature gives you the best of autofocus and the best of manual focus simultaneously. When you don’t have time to focus, you can let the camera do the job for you, and hope for the best. If you got a extra second or two, you can twist the focusing ring, tweak the focus, and get the shot.

Unless you want to focus the lens once and then take several shots with the same focus setting, or focus once and then wait for something to happen, I can’t see any reason to use the manual focus menu selection with the Sony lenses. When I use autofocus, I always use the mode that allows me to set where in the frame the camera should focus; when you twist the focusing ring to shift into manual mode, the place you set is the place that’s magnified. You don’t get a chance to go back and look at the whole frame before you take the picture, but you get used to that, and it works well and most circumstances.

Here’s a case of the camera refusing to focus on the branch because it’s too small in the field:

Here’s what I wanted:

Here’s the same thing with a differnt subject. What I got with autofocus:

What I wanted:

Next, an example with a small autofocus error. First, the whole picture:

Here’s what autofocus gave me, at pixel-for-pixel magnification:

And here’s what I got with manual focusing. Maybe this isn’t really an autofocus error per se; the AF couldn’t guess what part of the lichen I wanted in focus.