Is it possible to create a depth map using the EOS 70D Dual pixel AF?

Question

I've ordered the 70D and while waiting I had a though about the Dual pixel AF sensor.

My question if half out of curiosity and half out of it feeling like it might be useful.

I'm thinking that using the techniques used for the 70D AF one could perhaps create a depth map of the actual subject and assign approximate distance values to different parts of the image using the lens information.

(Creation of this thing is what I'm talking about: http://blog.topazlabs.com/wp-content/uploads/2014/03/depth-map.png)

So my question then, is it possible to create a depth map using the EOS 70D Dual pixel AF? Also, would this even be possible using a DP-AF type sensor and lens information?

(I'm assuming the answer to be "No" because while potentially it is theoretically possible the 70D is never going to forward the DP-AF information)

Answer 1

Possible with that kind of sensor? Yes. I've mused on that in another post here, about feeding it to software that simulates shallow depth of field.

Any small run of pixels that cross an image feature, read as two arrays of split pixels, will give phase information that is normally used to give an instantaneous focus command to go directly to the correct distance.

A single (split) pixel does not have depth. A short line of such pixels will be different on each split, shifted. So you need a long enough line to determine that and realize that it's blurry so you don't see sharp edges but a smear.

If you could save all of them to a raw file and process it on a computer, you could analyze it in a similar way to autofocus, spotting different spots and different depths. Then, expand around those readings to find an area with the same shift, or a smoothly changing shift.

Use that in conjunction with concentional image recognition to decide on the exact edges of things.

In a few years when 70D's and 7Dii's are appearing as junk, I expect hacking sites to publish how to play with the sensor chip pulled from an old body, with an Edison, π, or Arduino or somesuch, reading out and commanding it. For now, I'm not taking mine apart.

Answer 2

Pretty much every image captures depth information. After all, you can determine foreground and background, how far things are away, etc. You get a perception of depth when looking at a 2D image. This can be fooled by optical illusions.

There are algorithms that can do the same: figuring out the 3D model from a single 2D image: http://make3d.cs.cornell.edu/ Here's a bit about the background: http://cs.stanford.edu/people/asaxena/reconstruction3d/

I don't know if this could be improved by the additional information from the DPAF. Additional data never hurts. If it becomes a burden one can always ignore parts of it to reduce the amount. Your best chance of getting access to that data is probably to use Magic Lantern. I don't know to what extend they have access to it, but if anybody has, it's them.