Originally posted by Anvh I can not see how CDAF would be better in PDAF in term of speed in normal condition, the problem is that PDAF can see if the focus is behind or infront of the subject and adjust accordingly, it can even see how much it's out.
CDAF is just adjusting the focus and move the way that increase the contrast and that is simply trail and error.
That's clearly the reason why people believe PDAF to be superior by principle.
However ...
It is always worth to have some framework to make such judgements.
Both, CDAF and PDAF are cross-correlation maximization algorithms.
In the case of CDAF, it is the auto-correlation of a 2-dimensional region of interest with itself which is maximized.
In the case of PDAF, it is the cross-correlation of a 1-dimensional region of interest (narrow stripe) taken from two separate patches of the lens' exit pupil.
The PDAF case has the known advantage that the free parameter (the shift) can be optimized from taking a single measurement, while CDAF needs a sequence of measurements (3 at least). Say, CDAF needs a factor P more measurements wrt PDAF.
The CDAF case has the less known advantage that it use an order of magnitude more light, say, the amount of light it can use is a factor C over PDAF.
Now, if C>P, then CDAF outperforms PDAF in speed, if the mechanical focus latency can be made small (note that focus travel is the same for both methods).
My claim is that C>P will soon be achieved.
I back up my claim by experience I gathered from the development of a high speed CDAF for an industrial customer. I use an optimization function which allows to estimate direction and amount of shift from just two measurements made in immediate succession. The details are my intellectual property, so I cannot share.
As soon as CDAF outperforms PDAF (i.e., C>P), further advantages come to fruit:
- the measurement and focus travel can happen in parallel, speeding up the operation.
- subsequent measurements can correlate the region of interest first (focus tracking) which leads to a more reliable focus performance with long lenses.
- accuracy is a good as needed, no calibration problems, color shift etc.
- High speed AF.C bursts (24+ fps) may track the focus without further measurements at all, just using the taken photos as measurement. Also needed for video-AF.