General Optical Low Pass Filter swap project (to reduce moire and aliasing)

And don't forget the other big mystery, why does the 7D have so much color in the Moire and the GH1 does not.
 
grimepoch said:
And don't forget the other big mystery, why does the 7D have so much color in the Moire and the GH1 does not.
Click to expand...
It's not just the GH1; the charts I've seen from the Nikon and Pentax cameras don't show that color contamination either. It's something unique to the Canons (and they all have it; 7D, 5D, t2i, and 1D Mark IV).
 
Are they not skipping equally and using some strange algorithms to interpolate the coloring around? I'd imagine this is really a problem with their bayer pattern after the down sample.

or, from the fact we see different properties in horizontal and vertical aliasing/moire, maybe it's an artifact of using less data in one direction and not the other.

For instance, what if for the rows they measure every pixel or at least more pixels and use that data to down sample intelligently, yet for the columns, because they actually skip whole lines without reading them out or analyzing them, they get a different level of aliasing in both directions.

I guess looking at this one would want to measure the resolution in both the vertical and horizontal direction in terms of a resolution chart and see where each falls off and compare that difference first. Also, do this with a GH1 (or any other video camera) and then compare the two in both directions. This might start shedding some light on this.
 
Okay :)

So, this is pretty damn technical, but have a look at this comparison on canon/sigma sensors:

http://ixbtlabs.com/articles2/sigma9sd/

The reasoning for the coloring is given as:

"the Canon allows for a color moire for thin grooves, because of the different frequency of the elements sensitive to the different spectral distribution"

I'll be honest, I'm not sure how to interpret that sentence :) I *think* it is saying that the sensors do not measure R,G,B photons with the same sensitivity since the same exact photosite would measure them differently (as each photon of a different wavelength is carrying different energy). Do other manufacturers tune their pixels for the color they will be used for with their microlens/filter.

The Sigma used a Fovean matrix which I have researched before that doesn't require color filters. Not sure the pros/cons to this method.
 
Grime, it is much more simple than you try to make it be... math won't help anyone, ACTION will.

The only way to find out of the 7D can be cleaned up is to organize enough people to pony up portions of the OLPF custom built, then have one made to match the 7D's CLEAN resolved detail perfectly.

Then wait and see if it works...
 
Actually, math will save you money if you don't have unlimited amounts to spend :) I certainly don't feel it's worth ponying up 100 x 100 which is $10,000 for a test that at least doesn't have some mathematics behind it to figure out what you need in the first place. And that's assuming you hit it first target. Given a non mathematical approach is shooting blind, 3 tests could be $30,000+

Sorry, but that doesn't sound like a good return on investment. Like I said earlier, I fear that the right OLPF is going to reduce the beauty that people love with the 7D to a point it's not going to be desirable to shoot with.

The best analogy for you I have is this. When people master music, they typically start with the music at a higher bit depth than the target, in many cases, like for CD, you record at 24-bit but then release with 16-bit. The interesting thing is that before this reduction is done, they actually add shaped noise to the 24-bit audio before reduction. This gives an enhanced perceived dynamic range of I believe 21-bits of audio, even though it is not real range. (If you are curious, read about POW-R curves, this is VERY commonly used technology).

The apply this to video technology, think of a sensor that can only detect pure black or pure white. Pictures would look horrible right? Now, instead, allow the use of patterns, like half tones, in the image. You are still using just black and just white, but the patterns allow you to perceive a different level of detail by using shaped noise to represent those in between colors.

I believe that this is a similar phenomenon to what we gain in benefit from the 7D. The false detail adds to the perceived detail of the image many times in a positive way which people for the most part enjoy with the 7D (as we've seen with people's feelings on it).

How can you expect to resolve a problem without understanding the causes of it in a system that requires parameters to be selected when building a solution?

1) Someone first needs to determine what the 7D is doing with the sensor. How many rows are they skipping, what is the active pixel size, what are they doing with adjacent pixels on the same row. You can't design an OLPF that is basically duplicating an image at an offset in a different direction if you don't know how far it needs to move in all directions can you?

2) Even if you get rid of the moire, aliasing, who knows what other problems are going to crop up? What if the density of the active pixels used in calculations are different in X & Y and they employ different methods to interpolate the data? Which seems possible given Horizontal and Vertical performance of the resultant images are different.

Believe me, I would love to get rid of these problems that I have, but before spending a bunch of money, I think it makes a lot more sense to determine the actual cause of the problems.
 
Martti Ekstrand said:
It's not reserved for just Canon's HDSLRs either, check the chart on this review of the Vixia HF11 camcorder.

http://provideocoalition.com/index.php/awilt/story/review_canon_vixia_hf11_avchd_camcorder/P1/
Click to expand...

Which is what makes me most concerned that what one might think as the perfect OLPF for the 7D might not be. I mean, the HF11 is shooting at its optimum max resolution which one must assume the OLPF is specifically chosen for that resolution / pixel coverage and yet, there exists still a rainbow.

Don't forget too, which I am glad this mentions, is this:

"The best resolution we should expect from a 1080-line Bayer-pattern sensor is 80%"

Why does this matter?

If you skip two lines, if you look at individual colors, it means you have a green row every third row, but you get a red or blue row twice that since they are half as frequent. You have two different spatial frequencies in bayer patterns.

Gets even more complex then :)
 
Simple solution is to find a supplier for the birefringent material and cut it to size myself. That should be immensely cheaper than a full blown custom OLPF. I can start testing with just one layer "filtering" vertically with the horizontals un filtered.
 
Grime... all you need is a resolution chart, count the horizontal likes resolved accurately and then the vertical lines, and request an OLPF that cuts off just below that. Or you can keep talking till nobody reads your posts, again.

Shoca, hope you ca figure something out, trial and error with minimal expense sounds like a good way to go. Maybe you bump into the right stuff. :crossing fingers:
 
USLatin said:
Grime... all you need is a resolution chart, count the horizontal likes resolved accurately and then the vertical lines, and request an OLPF that cuts off just below that. Or you can keep talking till nobody reads your posts, again.
Click to expand...

You know, I was just trying to help and provide some analytical feedback and assistance for helping to make this work. Clearly, you'd rather approach it from an ignorant standpoint of the mathematics. Just because you can't understand it doesn't mean it isn't valid or necessary to understand if you are MAKING YOUR OWN OLPF. He already said that the cost is too prohibitive to buy one made for this sensor. When you order birefringent material, the thickness is important as it determines how far that second image is going to get projected on the image plane. Which is why they use multiple layers and rotate some of the layers to get particular patterns with the multiple images.

And just to correct you, it's not just the resolution on the chart, it also involves the size of the sensor in determining the OLPF.

But whatever, point taken, I've unsubscribed from this thread.
 
Last edited:
I guess I showed my frustration a bit too much, and I apologize if I offended you.

The fact of the matter is that Scarlet is around the corner and a fix needs to come in yesterday. I see it as a simple deal... you can't allow your sensor to get detail beyond X frequency (lines). Beyond that, it should be up to the shop that claims they can succeed, one that is willing to at least share in the R&D loss if it doesn't. Where you saying anywhere in your posts that you where such a shop? Do you have hands on experience?

Is there a shop that can make a custom OLPF?
How long would it take to be ready?
How much will the first cost?
How many of us would want one and how much would all of them + the first cost?

If we stray away from these types of questions or try to do it ourselves it will only add more parts to keep track of by everyone involved and it will NEVER happen. It is likely to never happen just from the fact that money needs to be raised for everyone's modification.

Again, sorry if I came off as a bit of an asshole, but I don't want to talk about it, I want to make it will HAPPEN, and find out if I can afford it.

Beyond that, if anyone is experimenting with materials in between the lens and the built in OLPF, an other very pragmatic approach, then I would L-OV-E to hear about that... cutting up a caprock and taping somewhere is more likely to help than trying to take a course on designing OLPFs... Scarlet is around the corner and this 7D becomes a still camera thereafter to people as concerned with Moire as those considering an OLPF replacement.
 
I did unsubscribe, but then, I am still interesting in the subject. I understand your frustration. I have money ready for the Scarlet now, and have been impatiently waiting as the release slips over and over (I see you over at reduser as well, I am Rick Burnett there).

I am a chip designer, not an optical engineer. I am fascinated by optics though, which is why I get so in to them :) There may be other resources out there that we've not found yet that can do some of this stuff cheaper. I also worry that the at home tests are very dangerous with regards to the cmos sensor. That shop already damaged the T2i and they do this professionally!

You are probably on to something with taking the caprock concept and trying it behind the lens. That would be a lot safer and easier to test.

In any case, no harm, no foul :) Like I said, I feel your frustration. Nothing worse than beautiful shots ruined by Moire I didn't see on the LCD.
 
Here's some quick math for filter properties:

Sensor size = 22.3 x 14.9 mm
18Mp (7D/T2i) Native resolution = 5184 x 3456
18Mp (7D/T2i) Native horizontal pixel pitch = 22.3mm/5184pixels = 4.3um
18Mp (7D/T2i) Native vertical pixel pitch = 14.9mm/3456pixels = 4.3um (hey, square pixels!*)

3Mp (D30) Native resolution = 2160 x 1440
3Mp (D30) Native horizontal pixel pitch = 22.3mm/2160pixels = 10.3um
3Mp (D30) Native vertical pixel pitch = 14.9mm/1440pixels = 10.3um

*Since 16:9 video capture is not the same aspect of the sensor, we still need full sensor numbers for the OLPF which are the same as the video resolution.

1080p native = 1920 x 1080
1080p new height in Canon 3:2 ratio = 1920 x (3456/5184) = 1280
1080p OLPF 3:2 full res = 1920 x 1280
1080p horizontal pixel pitch = 22.3mm/1920 pixels = 11.6um
1080p vertical pixel pitch = 14.9mm/1280pixels = 11.6um

720p native = 1280 x 720
720p new height in Canon 3:2 ratio = 1280 x (3456/5184) = 853
720p OLPF 3:2 full res = 1280 x 853
720p horizontal pixel pitch = 22.3mm/1280 pixels = 17.4um
720p vertical pixel pitch = 14.9mm/853 pixels = 17.4um


So with those numbers, it should be easier to ask for birefringent filter by spec. So if you wanted to blur vertically for 1080p on an APS-C sized sensor, you would need a filter suited for 11.6 micron pixels which is huge by current and past industry standards. For 720p, its a 17.4 micron pixel. Now to find a place that sells these things uncut.





*sarcasm
 
Last edited:
So, we know that our 1080p best case pixel size is 11.6um for this sensor size, and that our actual pixel in that area is 4.3um in a single direction, this means at a minimum, we need a birefringent filter that shifts the image out of that 7.3um of area that nothing is getting sampled.

Shaocaholica, so one thing I've noticed in in the specs on the birefringent material is that it basically splits the image into two and the second image is offset in a single direction. (Like the wiki post I put on here earlier). Using additional layers like I think you were planning will continue to duplicate the duplicates and essentially blur the image more which is needed.

The best way I can describe the issue is this, let's say you have a vertical line hitting the sensor through the lens in that 7.3um of dead space where the pixels are ignored on the sensor. The goal is to spread that image detail to surrounding pixels that are getting sampled. With one birefringent filter you will create a duplicate of that line at an offset. Since that line can be anywhere in that dead area, a single fliter isn't going to be able to do this because of the distance.

However, given my conversation with a company that makes OLPFs, what it appears you want to do is pick an offset difference relative to the pitch of your original pixels. so if the pixels are 4.3um, then you can use that as a shift distance and then use as many layers as skipped pixels + 1. that way you are guaranteed that the detail is merged together by the duplicates up to the level of resolution you can record with the sensor at this level.

So in this case, if we assume 2 rows are skipped, then we use three layers that each duplicate at an offset of 4.3um.
 
No no, 3 birefringent filters. OLPFs are made of multiple layers of birefringent filters and not all of them in the same direction. So say you look at a singular beam of light, they pass through one and you get two beams now, the second let's say offset directly above it at a distance the thickness was tuned for (where that second dot is depends on the orientation of how the filter is rotated).

Now, you pass that through another filter that is rotated 90 degrees clockwise from the first, now you have 4 dots in a square with two on the right duplicating from the left. These are the types of things they are doing in the OLPF the engineer told me.

I extrapolated backwards that we need to take 11.6um of spacial information and get it into that 4.3um actual pixel size without missing it. A single filter would not work. You are limited by your spacial sampling size. So with 4.3um and multiple layers, you make sure that everything across that dead area on the sensor is represented in the sampled 4.3um space by duplicating it at a 4.3um pitch with the filters. That's pretty damn cool they came up with that method! :)

As a side note, since the pixel is actually smaller than 4.3um, most CMOS sensors incorporate a microlens with the color filter to pull the light in from the inactive area of the pixels so that area is not dead.
 
grimepoch said:
No no, 3 birefringent filters. OLPFs are made of multiple layers of birefringent filters and not all of them in the same direction.
Click to expand...

I get that part. I was just excited for a moment that a solution might be viable using already available parts.

What we do have right now is D30 OLPFs. I wonder if stacking 2 might produce something usable. I don't know if already assembled OLPFs can be easily deconstructed nto the original layers so the easiest thing to do is to just stack them.
 
You must log in or register to reply here.
Back
Top