C200: Helping DPAF

R

rob norton

Veteran
Sometimes face detect doesn't work in low light conditions. Can you overexpose on purpose to help it out? I'll be shooting a low light, RGB filled music video soon and am wondering if this is a possibility, specifically when shooting raw?
 
Why not do a test and see how far you can overexpose and still get acceptable results in post?
 
My guess is that it would make no difference.

The DPAF uses light that hits the sensor. Whatever camera settings you have (apart from iris and ND) dont actually affect the light that hits the sensor.

WHat you dont want to do is to lower the light hitting the sensor by using a small iris (in a low light environment)

So
shoot the widest iris you feels delivers an acceptable look
dont use ND (of course!)

if the footage looks too bright pull it down in post.

A narrow DOF also helps give the AF something to 'bite' as the focus is dropping off fast.
 
I was shooting a reality tv show a few months back and the scene was at a wedding with DJ and low light...(music video feel). I had issues with DPAF for sure. I wasnt using face track as that does not work with the Sigma 18-35 1.8. Either way it struggled. My Canon 24-70 II struggled as well. On the other hand my Canon EOS R performed much better in lowlight with the latest DPAF tech. It’s my favorite Gimbal camera. I can’t wait for an RF Mount Cinema Camera priced around the C200. Dying to use the 24-70 F2 :p
 
My impression (I have not tested this theory thoroughly) is that the simpler, center-of-the-sensor DPAF on the C100s works great in almost any lighting condition, but the more advanced DPAF on the C300mkII and the C200 really does require decent light hitting the sensor to work well.
 
drummondb said:
My impression (I have not tested this theory thoroughly) is that the simpler, center-of-the-sensor DPAF on the C100s works great in almost any lighting condition, but the more advanced DPAF on the C300mkII and the C200 really does require decent light hitting the sensor to work well.
Click to expand...

I concur. I used to have the C100 MKI with DPAF and it worked very reliably whereas the DPAF and face detect on the C200/C300 MKII are definitely more finicky.
 
The DPAF doesn't work almost at all on anything below about 20% IRE. This is where DPAF and continuous auto focus on the camera falls apart and starts hunting. I always keep my manual focus guides on because the manual focus guides will turn gray when the DPAF doesn't where to focus. If you turn on one touch auto focus or continuous auto focus when the manual focus guides are gray, the camera might just start hunting for focus and go all the way to infinity or close focus.

Face tracking doesn't work in low light or high contrast very well. I'm not really sure what it is looking for, but I am guessing that it is a mix of skintones and high and low levels indicating eyes. If one of the eyes is in the dark, it doesn't seem to see a face. Face detection really doesn't like glasses or dark faces. With high contrast, I would just use regular tracking as the tracking will latch onto whatever high contrast area is created by the light hitting the person's head. Mix that with face priority for when the person's face is fully lit and a face is recognized. There is no reason to underexpose with RAW. ETTR is a good policy with any RAW image.
 
I like DPAF and have at times have been a power user of face detect. Using face detect in continuos mode with variable lighting can be risky. The things that I have found that spoof face detect are Glasses, especially shades, and in combination glasses (shades) with beards, hats and low light. If shooting one of the log profiles, having view assist turned on, and a lut applied to the VF or LCD seems to help ( maybe increased contrast is the key) .
The best thing I have come up with is mapping the button on the grip to AF lock, if in AF continous mode, or some what different- map the grip button to one shot. In either case you get some benefits of DPAF with the ability to toggle it as needed.
 
It might help to understand how face detection works: https://www.pyimagesearch.com/2015/...-for-object-detection-with-python-and-opencv/ (on a camera it's implemented in C/C++ and/or custom hardware (and/or GPU-like array processors- e.g. Intel Movidius as used in drones)).

The 'face box' sub image is processed and (effectively) multiplied against a kind of mask (e.g. Eigenface: https://en.wikipedia.org/wiki/Eigenface ) to get a score, and when the score passes a threshold, reports a face has been found (depending on the implementation, it's a bit more complicated, but the idea is the same: pattern recognition via weighted sums of values). So you can see that if the input image is too dark, too light, too much or too little contrast, face is altered (glasses, beard, hat, hair-on-face, extreme shape), or too close too far, the score will be too low and no face found.

For those with a background in video and stills compression and the DCT (discrete cosine transform), you'll find the Viola-Jones AdaBoost method to look familiar as summed frequency component classifiers: https://cs.gmu.edu/~kosecka/cs682/lect-face-detection.pdf .

Basic face recognition (FR) is pretty old in the AI/machine learning world, and there's even decent implementations in Javascript! https://medium.com/@muehler.v/node-...ce-recognition-using-deep-learning-ea5ba8e852 . For the coders out there (or not), Javascript is a relatively slow, interpreted language that runs in a browser (and now on the backend e.g. node.js). Also note that once a face has been found, finding the eyes, nose, mouth is much easier, because it's known approximately where those features typically exist. From my tests, Canon knows where the eyes are, just doesn't draw them.

In 2019 we have ResNet et al "Deep Face" deep learning based methods (from MS, Facebook, Baidu etc.) which are vastly more powerful/accurate: https://arxiv.org/pdf/1804.06655.pdf . Sony will soon release advanced AF / FR not just for humans, but also for animals! That's a much bigger database, yet still not a big deal in 2019 in terms of memory and computing power, even for a camera/drone/embedded system.

The Sony A6400 is supposed to have this new AF system, and the A7 III and A9 will get an upgrade around April. From the demos online, this puts Sony way ahead of everyone else for AF / FR, and it's going to get much better, even in low light, with glasses etc. and other problem cases.

Canon could release firmware updates to fail more gracefully with AF, and possibly even improve quality (not clear how much of their system is programmable vs. fixed functions in hardware ASICs). If not their next gen cameras will certainly have updated hardware to compete with Sony (and even Fuji and Nikon: not having good FR is only a cost reduction issue- the tech is well known now for pretty good results). When adding infrared and 3D sensors, the quality gets even better (e.g. for latest iPhones as a security feature).

In the near future, I'd expect camera companies to allow you to train the database to recognize specific people (and animals and other objects!), and even program the camera to do focus pulls between known faces/objects. All easily doable already in 2019.
 
You must log in or register to reply here.
Back
Top