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

Barry_Green said:
I'd help out.
Click to expand...

You are helping Barry:happy:. Lets get the side by side done and publish the results even though we know the gist of what to expect. Hopefully that might stir up some internet ramblings beyond DVXUser and get the attention of Canon and pals that people are trying really hard to fix this stuff on their own.
 
Two things:

a. Told you so. He he he... You need to cut detail at about 720p. :thumbsup:
b. We won't know how much of an improvement is there till you can do a side by side.

(if you want to shoot something super fast before shipping it out to Barry I am up for it)

c. You da' man for doing this!
 
USLatin said:
Two things:
a. Told you so. He he he... You need to cut detail at about 720p. :thumbsup:
Click to expand...

Cutting detail at 720 will yield the same results. I'm willing to bet. This is primarily a readout problem rather than a OLPF problem.

Try looking at something through horizontal blinds. This is how the camera sees the world in video - it sort of interpolates the missing information that's not being read out. This will result in gapping/warping/aliasing/moire etc.

Once the cameras start downsampling all 18mp (or whatever number) in real time to 1080p we will probably see an improvement. Right now it's subsampling instead of downsampling and that is where I am betting most of our gremlins lie.:bath:
 
Thanks for the offer but I literally just handed the camera off to Philip Bloom who's taking it to SXSW to give to Barry.
 
Dan_Kanes said:
Cutting detail at 720 will yield the same results. I'm willing to bet. This is primarily a readout problem rather than a OLPF problem.

Try looking at something through horizontal blinds. This is how the camera sees the world in video - it sort of interpolates the missing information that's not being read out. This will result in gapping/warping/aliasing/moire etc.

Once the cameras start downsampling all 18mp (or whatever number) in real time to 1080p we will probably see an improvement. Right now it's subsampling instead of downsampling and that is where I am betting most of our gremlins lie.:bath:
Click to expand...

A stronger OLPF acts like an optical downsampler not just a blur filter. It should work in theory if the right srength is used. The problem is that these things can't be bought cheapy at the strengths we want, even just to test.
 
The problem is, if we go to an OLPF for 720p resolution, the sharpness is going to drop considerably I believe, which was the big question even at this OLPF. Dan, I think the OLPF for 720p will make the Moire and Aliasing disappear, but I think the resultant image is going to have a lot of blur in it for 1080p mode. It probably means you'll need to shoot in 1080p and then in post down convert to 720p to get the best image (which is what i do now for a lot of the projects we do).

Two things I have wondered:

1) For each row, do they shift out the entire row and then down sample the row giving us better horizontal image?

2) Since they skip SO MANY lines, there is a lot of dead space not being used. I kind of agree that since the ratio is not even on what it skips, I bet this is why the actual resolved image is so poor in terms of how much we'd need to blur the image.

Interesting though, and I too would actually donate a little money to you for doing this test.
 
Sharpness will not drop. Perceived sharpness perhaps, but that is all manufactured non-existent stuff... and yes, this would be for 720p downconversion. There IS no 1080p in these cameras. We are talking about getting what they actually have to offer without all the smelly rotten growth all over it.
 
Yes, I see what you are saying there and I agree :) It's sharp when it lines up perfect with the rows that are being used. Hehe.

And yep, downconversion is the way to go. Unfortunately.
 
Yes, but not really a misfortune when you consider that for this much $$ you get a S35 sensor with quite a bit of dynamic range and great color! (thank you RED for changing the market)
 
Well except when you cannot see that ugly moire till you pull it off the camera because you don't have a 720p ref monitor with you (although I am seriously thinking of getting the smallHD monitor).

No, it's a great little camera in seriousness, I just get frustrated when I look through shots and see a bunch ruined when shooting urban settings and I was a little too deep on the focus.

But I agree, thank you RED for changing the market. Now if they'll just get them out the door!!!! :)
 
I know. Though I was talking from the point of view of a (theoretical) perfectly modified 7D with a ~720p OLPF.

I want a smallHD monitor, bad. It looks like a great buy, and many recommend it.
 
So is it worth trying a 720p OLPF? I'm not sure if it is at this point.

The smallHD people are right here in the town I live in. Gonna go visit them probably next week.
 
grimepoch said:
So is it worth trying a 720p OLPF? I'm not sure if it is at this point.

The smallHD people are right here in the town I live in. Gonna go visit them probably next week.
Click to expand...

You sound as if you could get one pretty easily:happy: A 720p OLPF at APSC-C size is practically unheard of. Same thing for 1080p.

Seriously though, I'm still trying to source a stronger OLPF to try but obviously it can't cost a fortune. Im looking at the raw materials rather than the completed units.
 
shaocaholica said:
You sound as if you could get one pretty easily:happy: A 720p OLPF at APSC-C size is practically unheard of. Same thing for 1080p.

Seriously though, I'm still trying to source a stronger OLPF to try but obviously it can't cost a fortune. Im looking at the raw materials rather than the completed units.
Click to expand...

Haha! I meant in terms of results, not in feasibility. From a mathematical standpoint, I'm wondering if it's worth it because I cannot find anything with regards to the ratio of active area to non-active area in OLPF design. There has to be a serious drawback in terms of having such a small row related to how many rows are skipped.


X-------X-------X-------X-------X

In this example, the dashes represent seven skipped rows. 14% of the area is used for active pixel area. 86% of the area is wasted. That means that detail in that 86% of the area has to be spread across that large space to the neighboring pixels to prevent the aliasing/moire.

A sensor that has larger active pixels so that the sensor size is the same size, let's say like
XXXXX--XXXXX--XXXXX--XXXXX--XXXXXX

(sorry the X and - are not the same size)

In this case, the dead area between is much smaller. Resolution is the same, but the second sensor already picks up the light since it has less dead space between.

When you are selecting the OLPF, how are you deciding what it is that you want? I think the metric needs to be selected that would be diffusing the light from 1/2 the distance between active pixel edges minimum on the projected image that passes through to the sensor plus half the width of the active pixel. For instance (and these are theoretical numbers) if the dead space was 50um, then you'd want to spread the light out at minimum of 1/2 50um + 1/2 the size of an active pixel.
 
I'm not at all sure that OLPC manufacturers actually control where there are "hot spots" or not, I would think it was just an overall blur, since if they had that kind of control, they would have to match the filter precisely to the pixel layout of the imager. One thing that concerns me is that the resolution may actually be much lower, say 1/2 720P if it really is one hot pixel per 9 raw pixels, which is hard for me to believe... What about asking one of these types of custom manufacturers to see about supplying you with a one-off sample?

http://www.optics-online.com/lpf.asp

Ahh, never mind, I see what you're saying. Perhaps with the correct charts, we could figure out the real resolution delivered in video mode...
 
doccutter said:
I'm not at all sure that OLPC manufacturers actually control where there are "hot spots" or not, I would think it was just an overall blur, since if they had that kind of control, they would have to match the filter precisely to the pixel layout of the imager. One thing that concerns me is that the resolution may actually be much lower, say 1/2 720P if it really is one hot pixel per 9 raw pixels, which is hard for me to believe... What about asking one of these types of custom manufacturers to see about supplying you with a one-off sample?

http://www.optics-online.com/lpf.asp

Ahh, never mind, I see what you're saying. Perhaps with the correct charts, we could figure out the real resolution delivered in video mode...
Click to expand...

I already asked these guys and they said the size and spec were so odd that it would have to be a special order with a 100 unit minimum at $100 each.
 
Here is a good example if anyone is wondering what birefringence is:

http://en.wikipedia.org/wiki/Birefringence

OLPC manufacturers will control the offset of the split. Remember, photons are coming in everywhere, so they use multiple levels of birefringent material to split that light in multiple directions. The distance is probably a function of pixel size versus dead space.

The row skipping I mentioned was not calculated by me, I think some people reverse determined this from test chart images. Someone else will have to chime in there. At best case scenario,
Max image size for the 7D is 5184 x 3456, so thinking vertically, you'd have to skip two lines to get down to 1152 lines and then throw some out to get to 1080. Also, you'd need to skip actually an even amount to keep the bayer pattern correct:

GRGRGR
BGBGBG - skip
GRGRGR - skip
BGBGBG

The problem is, shaocaholica has to determine just how far that image needs to be moved X & Y with birefringent optics to get past the dead space and into the next pixel area, like the math I posted. I can DEFINITELY see why these aren't common sizes. It depends on the pitch of the pixels, the size of the pixels, and the ratio to active space compared to dead space.
 
Last edited:
You must log in or register to reply here.
Back
Top