Canon 70D Dual Pixel CMOS AF

A Close Look at the Canon 70D's Dual Pixel CMOS AF

by Dave Etchells

As the years go by, we see camera technology advance by fits and starts. Some developments are a bigger deal than others, but it's rare that anything really amounts to a true technological breakthrough. However, the Canon 70D's Dual Pixel CMOS AF system qualifies.

This new AF system is as revolutionary a development as any we can remember seeing since the dawn of the DSLR era itself. By integrating accurate, fully-capable phase-detect autofocus over a majority of the image sensor's surface, Canon is fundamentally rewriting the book on autofocus.

The Canon 70D delivers phase-detect autofocus across an area that's fully 80% of the height and width of the sensor, that allows any area within that region to become a focus point, that can remain operational during video recording, and that will operate at any aperture. It's a whole new AF ballgame, and one that's going to shake the DSLR video business to its roots.

So what's all this Phase-Detect stuff about, anyway? Let's start by understanding how camera autofocus systems work. There are two main types of autofocus, namely contrast-detect and phase-detect. Consumer digicams invariably use the former, while SLRs use the latter. Each has its advantages, but phase-detect systems are generally faster, and they have other benefits as well.

So-called contrast-detect AF works by analyzing the image to see how sharp details are, and then trying to adjust the focus to increase sharpness. The reason it's called contrast-detect is because the sharpness measurement consists of calculating the contrast or brightness difference between adjacent pixels across the AF frame. As the image becomes more sharply focused, this contrast increases, so focusing is just a matter of maximizing the overall contrast of the subject.

The thing is, the only way to tell whether the subject is in focus -- determining that the contrast is at a maximum -- is to nudge the lens one way or the other, and see if the contrast increases or decreases. This takes time, and if you're recording video, you'll see the image shift in and out of focus as the camera "hunts" for the best setting.

By contrast (sorry, bad pun), phase-detect AF works by actually measuring how much out of focus the subject is, and in which direction, then moving the lens exactly the amount needed. This is much faster, and there's no hunting involved.

The trick, of course, is measuring the out-of-focus amount; how does the camera do that? If you're old enough to remember split-image viewfinders on film SLRs, that's exactly how phase-detect autofocus works, only electronically. The key is looking at light rays coming through opposite sides of the lens, and seeing whether they line up or not. If you think about it, for an object to be in focus, the light rays from both sides of the lens have to meet at the same place. If the subject is out of focus in one direction, the light rays will run into the sensor before they have the chance to converge. If it's out of focus the other way, the light rays will pass each other before hitting the sensor. The illustration at right (courtesy of Wikipedia) illustrates this.

In SLRs, the phase-detection measurements are made via a separate AF sensor, usually located in the bottom of the mirror box. A portion of the mirror is partially-silvered, letting light pass through, bounce off a secondary mirror, and then down onto the AF sensor itself. This brings with it some limitations.

For one, the AF sensor needs to be precisely positioned, so light rays will travel exactly the same distance to reach it as they will to reach the image sensor. If the AF sensor is just a little out of alignment one way or the other, the camera will back-focus or front-focus (that is, focus behind or in front of the subject). These days, many cameras let you make micro-focus adjustments electronically, but you often need different adjustments for different lenses, or even different focal lengths of the same zoom lens.

For video shooting, the huge disadvantage is that the phase-detect AF system can only operate when the mirror is down. Raise it and the light goes to the image sensor, so it's an either-or proposition. A few years ago, Sony developed a unique translucent-mirror technology used in their SLT cameras, which lets phase-detect AF operate all the time. This was quite a breakthrough, as it's not only better at tracking moving subjects (since it's "looking" at them all the time), but it can also remain active during video recording.

This leads us to another major limitation, though: Because it needs light rays approaching the sensor from different angles, phase-detect AF typically can only operate at fairly wide apertures, usually f/5.6 or larger. This isn't a too much of a problem for still photographers, as the lens can be kept wide open for focusing, then stopped down to take the actual picture. Shooters with long lenses and teleconverters can have a hard time, though, because that combo often results in maximum apertures that are too small. The aperture limitation also rears its head with Sony's SLT system: While the AF can remain active during video recording, you can't stop down the lens and also keep your AF going as well. If you want to shoot video with the lens stopped down, you'll be back to manual focusing again.

On-chip Phase-Detect AF. Recently, technology has been developed that places phase-detect elements directly on the image sensor itself. Pioneered by Fujifilm, but now available from a wide range of camera makers, this technology has to date been only a partial success, as many such systems have proven to be rather slow, and haven't entirely eliminated focus hunting. (The exception being the Nikon 1 series, which manage to focus quite quickly indeed.)

On-chip PD systems work by "shading" some of the pixels on the array, so they'll only see light arriving from one side of the lens or the other. With CMOS sensors, pixels can be read out individually, so the camera can use these special, shaded pixels as if they were part of a separate phase-detect sensor, to compute subject distances.

Most systems using on-chip PD technology are labeled "hybrid AF," because it turns out the phase-detect system can't get the entire job done on its own. Instead, what happens in most cameras is that the on-chip PD elements are used to get the lens close to correct focus, then contrast-detect is used to fine-tune it the rest of the way. This can cut some time relative to a purely contrast-detect focusing system, but it's not as fast as full phase-detect focusing, and there's still some hunting involved to achieve and maintain focus.

Another issue with on-chip PD as it's been practiced thus far, is that the pixels used as PD elements are shaded, meaning there isn't as much light hitting them as their neighbors, so the camera's image processor has extra work to do to compensate for this. It may not be a lot of extra work, but there's clearly at least some hit on the processor to cope with it.

Perhaps because of this processing overhead (or perhaps for other technical reasons as well), on-chip PD systems to date have involved only a relatively limited number of PD pixels, behaving more or less like the AF points on a conventional SLR's AF sensor: Frame coverage has been limited and discontinuous.

The breakthrough: Canon Dual Pixel CMOS AF. With the sensor used in the EOS 70D, Canon broke the mold, asking why it shouldn't be possible to equip every sensor pixel with phase-detect capability. Of course, in doing so, they couldn't arbitrarily throw away half of the light falling on the chip, by shading the pixels the way conventional on-chip phase-detect systems do. That would cost them a full stop of ISO sensitivity, and who knows what odd optical artifacts might result as well.

The solution was to split each sensor pixel into two sub-pixels, each receiving light from the half of the lens' light cone that it normally would in a sensor with conventional microlenses. I'm not enough of an optical scientist to be able to describe in detail how light travels through microlenses, but it seems reasonable that light rays striking a microlens from one side of the frame end up hitting the same side of the underlying photodiode. It's not clear at this point whether Canon figured out a way to insert a light barrier vertically into the microlens structure to improve the separation between light rays arriving from opposite directions, if they somehow altered the microlens structure to achieve the same end, or whether conventional microlens structures inherently create enough separation on their own. Whatever they did, it apparently does give them enough separation between light arriving from different directions that they can perform effective phase detection.

However they're manipulating the incoming light, they've split the photodiode for each pixel into two halves, so they can read out image data corresponding to top-arriving and bottom-arriving light rays separately. Based on what Canon has told us, they can also read out both sides simultaneously, for normal imaging. In fact, it appears that they can read out data from the two pixel halves separately and together at the same time, when they're recording video. That sounds a little tricky to do on a pixel-by-pixel basis, but perhaps they're taking advantage of the much lower resolution of video to do the two different types of readout on alternating rows of pixels.

We believe that the split pixels are still managing to collect the same amount of light overall as a conventional, un-split structure, based on a study of the Canon 70D's low light/high-ISO performance. Were they only collecting only a portion of the incoming light, we'd have expected that to show up as significantly decreased sensitivity/higher noise, and we didn't see any such increase. It's possible, though, that sensitivity / noise improvements in other areas of the sensor and imaging pipeline are sufficient to mask any deficit caused by the split-pixel design. Only Canon knows for sure if there's a drawback at the sensor level, but the important thing to note is that overall we saw slightly better high ISO performance from the 70D than its predecessor, while at the same time it offers the uniquebenefits of Dual Pixel CMOS AF.

The Canon 70D's Dual Pixel AF operation varies slightly, depending on whether it's in still shooting or video modes: When running in still capture mode, the camera determines the degree of focus adjustment needed before commanding the lens to move, while in video capture mode, focus determination and lens movement occur continuously and simultaneously.

Canon 70D Dual Pixel Autofocus Options

When it comes time to actually use the Canon 70D's Dual Pixel AF system, you have four options. They are:

• Flexi-zone with a small, movable box: This mode gives you a smallish box that you can move anywhere in the frame you want, at least within the 80% vertical and horizontal boundaries mentioned above.
  
  
• Flexi-zone with entire AF area: This mode displays large brackets on the LCD, delineating the 80% area coverage of the AF system. Anywhere within that area, the AF system will select and lock onto the subject that's closest to the camera.
  
  
• Face Detect mode: Nikon has for some time now taken advantage of its relatively high-resolution RGB exposure sensor to provide face detection, in conjunction with the camera's specific phase-detect AF points. The idea there is that the exposure sensor-based face detection selects the conventional AF point overlaying the detected face to set focus with. This was a very welcome feature, and Nikon deserves high praise for bringing it to market, but it of course was limited by the fixed locations of the AF points.
  
  With the Canon 70D, though, face detection has no limitation other than the inherent 80% area mentioned above. The system can precisely locate a face anywhere within that area, and place an AF box exactly around it. It's not a matter of picking the closest fixed point overlying the detected face, but rather constructing a custom "AF point" that exactly matches the face. This is huge, as you don't have to give a thought to whether your subject is near an AF point or not: You can create a custom AF point anywhere within the available area, and that AF point can move to follow the face. Very cool!
  
  
• Touch Select AF: This is especially nice for videographers. We applauded a similar feature on some of Panasonic's Micro Four Thirds cameras in recent history, that let you set focus points during live video recording simply by touching the appropriate point on the LCD screen. The touch-select AF feature on the Canon 70D works similarly, but with phase-detect rather than contrast-detect AF, so there's no hunting (although cameras like the Panasonic GH3 do remarkably little "hunting", given that they're using contrast-detect AF).
  
  Canon shared with us a movie called "Handmade" (below), which shows just how useful this feature is in video production: Also check out the "Making of Handmade" video (further below) for examples of just how well and smoothly this feature works. There's unfortunately no control over how quickly or slowly the focus racks between positions, but the rate Canon chose seems like a good compromise for most situations.

When used with one of Canon's new stepper-motor-equipped (STM) lenses, the resulting autofocus operation is not only quick and smooth, but supremely quiet. Even in a room with relatively little ambient noise, we've been hard-pressed to hear any focus-motor noise on the audio track.

Why is the Canon 70D's Dual Pixel Autofocus so cool?

As I mentioned above, we see the new Dual Pixel CMOS AF technology in the Canon 70D as a genuine game-changer for DSLR video. Why? Here are a few reasons:

• Fast, and no hunting. First and foremost, this new AF system is not only fast and accurate, but there's none of the "hunting" that's an unfortunately necessary side-effect of contrast-detect systems. The camera always knows exactly how much the lens is in or out of focus, so it doesn't have to rack focus back and forth to figure it out. Seeing the 70D's AF system working next to a conventional contrast-detect one makes us realize just how much we've been putting up with, just to be able to autofocus during live video recording. There's simply no comparison.

• No possibility of front- or back-focusing. As we discussed above, conventional phase-detect AF systems can misfocus if the AF sensor is slightly out of position, or even due to differences between lenses' internal calibration settings. This has always been the strong point for contrast-detect AF, namely that the main image sensor is being used to determine focus accuracy. When something's in focus with a contrast-detection based system, it's by-golly in focus.
  
  With the advent of full-sensor phase-detect AF in the Canon 70D, the same is now true for it as well: When the system says things are in focus, there's no question of whether a separate AF sensor is lined up properly with the image sensor; they're one and the same. (Of course, this is only true for Live View mode. In standard shooting, the 70D still uses a separate phase-detect sensor that's subject to the same tolerances as any other DSLR. And even with the on-chip phase-detect, there are always possibilities for error with any system. In our testing, we found the 70D's focusing to be very accurate.)

• Superb focus tracking of moving subjects during video recording. As you can see in our focus test published last August, focus tracking on the Canon 70D is superb. Bottom line, the camera behaves exactly as any consumer (or pro, for that matter) would expect it to: It held focus on the subject, without wavering.

• Extremely accurate. In any phase-detect AF system, focus accuracy is a strong function of the "baseline" of the AF sensor. The greater the angle of the incoming light rays and the longer the distance between AF sensor elements, the more precise the focus will be. This is why many higher-end cameras have special "high accuracy" central points, that kick in at apertures of f/2.8 or larger. Taking advantage of the wider light cone at very wide apertures, they use a second set of AF pixels spread further apart from each other to achieve greater accuracy.
  
  With the Canon 70D's image sensor, the effective AF baseline can be made as wide as the lens will allow, since there's no restriction caused by a fixed AF-sensor geometry. As the aperture gets wider (and you really need the added focus precision due to shallower depth of field), the camera can simply select a more widely-separated group of pixels to use for its focus measurement. There's no limit to this, either. An f/2 lens will give higher AF accuracy than an f/2.8 one, and an f/1.2 lens will be even more so.

• No aperture limitation. This is huge. Again as we described above, most photographers are accustomed to the fact that their SLR AF systems just won't work with apertures smaller than a particular limit, most commonly f/5.6, but sometimes as high as f/8. F/11 or f/16? Forget about it.
  
  Thanks to the ultra-flexible, full-array nature of the Canon 70D's AF system, it suffers no such limitation. This is great news for photographers using long lenses with teleconverters. Put a Canon 100-400mm f/4.5-5.6 zoom on a 2x teleconverter, and there's no chance it will autofocus on a normal SLR body, but the 70D will have no problem so long as you're shooting in Live View mode. Videographers will also be thrilled to be able to stop down while recording video, without losing AF. Contrast-detect based AF systems can of course work at any aperture setting, but at the cost of slower operation and focus hunting.

What won't the Canon 70D's Dual Pixel AF system work for? If all this sounds too good to be true, it probably is; nothing's ever a complete panacea. In the case of the 70D, Dual Pixel AF shouldn't be your first choice for applications like fast-action sports shooting. But don't worry, the Canon 70D inherited the same dedicated phase-detect AF system found in the EOS 7D, which is active when using the optical viewfinder. The system features 19 autofocus points, each of them a cross-type, optimized to detect both horizontal and vertical edges. In the center of the screen is the X-type sensor, designed to detect diagonal edges as well. It also requires lenses of f/2.8 or better, while the other points will work up to f/5.6. It won't bring any of the benefits of the Dual Pixel AF system, but it also won't require use of Live View mode. You can read more about it in our Canon 7D review.

Pro or Amateur: Who's it for? Sometimes, technological advances only make a difference for either pros or amateurs, but Canon's Dual Pixel CMOS AF is a fantastic innovation for everyone. For the professional videographer, it's a seriously big deal, for all the reasons we just discussed. For amateurs, it might be even bigger, though: Finally, there's a DSLR that shoots video the way the average consumer would expect it to.

In the full-area Flexi-Zone focusing mode, you can just point the camera at your subject, and it'll focus and stay focused, regardless of what the subject does, as long as there's nothing else closer to the camera. Or, just use the Touch-Select mode, and focusing is as simple as pointing your finger. When you're recording, there'll be no drifting in and out of focus, the picture will stay crisp and sharp. If pro videographers will flock to the Canon 70D, amateur users should even more so.

Bottom line, Canon's Dual Pixel CMOS AF is a true quantum-leap in technology. In fact, we may have already found our Technology of the Year for 2013.