Thailand & Taiwan & Singapore 1970's Super 8mm

[MattiasN]

Hi new film clip again some of the scenes have blown highlights

https://vimeo.com/82497785

[S8 Booster]

hi. on my screen it looks very good - like 16mm of the 1970s like projected.

shoot....

[Tommy]

Superb quality transfer, indistinguisable from 16mm......can you describe your type of transfer system?

[MattiasN]

Thanks

the movie was from ebay
i have 2 more reel´s left to convert this was reel number 2

Reel1: Hawaii 1970's. Asian american couple visiting the Isle. Sometimes features landscapes like in Lost, sometimes city skyscrapers, beaches, asian style gardens, and then a crusade in the port with militar ships moored, and many other sights from that decade.
Reel2: South East Asia (Thailand, Taiwan, Singapore), 1970's. Magic trip throungh Buddist temples, river crusade like in Hollywood movies we watched many times, rising cities (very different from now) and incredible nature.
Reel3: USA 1960's. This is an incredible document about the typical American family having trips to places like Universal studios theme park, Niagara Falls, National parks (featuring the amazing meeting with a Bear walking in the street close to the car). You will see strange looks like in Mad Men, typical cars and lot of things of the American history.


here is some photos of the reel´s he did link on ebay when i did buy it
http://s1082.photobucket.com/user/danie ... t=3&page=1

the transfer system is a home made you can see photos here http://www.flickr.com/photos/94271811@N03/
i run the film two times thru the projector one for low exposure and one for high exposure
then i use avisynth script to vertical stabilize sprocket hole on the low exposure and then align high exposure to it
then i feed the low and high exposure separately as Clip_A and Clip_B because now they are align in to the hdr function in avisynth and they come out as merge singel clip then i use autolevels function and last i use imagewriter in the avisynth script then it output it as image sequence

all that i do in one pass in avisynth but it goes very very slow about 0.52 fps it take some time this film clip was 14985 frames so it did take about 8 hours

next i do the denoise and sharpening with avisynth and the speed is the same so it take 8 hours to

[carllooper]

More great work Mattias.

That HDR sampling you do really works wonders. Kodachrome 40 is (was) also a superb film. In the HDR method you're using (double pass) it's not just an increase in dynamic range you'll be getting, but also an increase in resolution, due to the small offsets in the pixel alignment on each run. The more passes you do the greater the increase in resolution you'll get, but if merging the passes at the original capture dimensions that additional rez information is transformed across into the dynamic range domain.

I'd suggest experimenting with a larger target dimension (eg. 4K) as a master. And instead of doing just two passes, do four (or more). You expand each capture to 4K (which in and of itself doesn't do anything) but when merged (in 4K space) the overlapping pixels from each capture will interfere with each other and induce a higher resolution version of the transfer. One of the benefits here is that you can do some sharpening (and other fine tunings) in 4K which, when the image is subsequently downscaled to 2K (or less), looks a whole lot better than if one had otherwise done the fine tuning in 2K (or less).

I'd say its not a way of making Super8 look like it was 16mm but a way of making Super8 look closer to the way it should look in the first place!

In any case well done. I'm very impressed.

C

ps. I've been using a Canon DSLR in my own work but am thinking of getting the 4K version of these (under $1K):

http://ww2.ptgrey.com/USB3/Flea3

At 4K it has a data rate of 21 fps. At 3.5 K (ie. transferring a sub-region of the sensor buffer) it'll run at 24 fps. My Canon, while 5K, runs at something like 1 fps when doing on the run RAW transfers.

For location work, you could combine this camera with a tablet PC, in a wooden box (or whatever), and you'd have your own custom 4K (or 3.5K) digital movie camera.

[MattiasN]

Thanks

the film was even good without the denoise but i did want it little more sharpen then i have to denoise little
now when i do the denoise i overlay the original clip with the denoised so i get the grain back and some of detail that get lost when i denoise

i my self have checked at this http://grasshopper3.ptgrey.com/USB3/Grasshopper3

or if i did have to much money

http://www.edmundoptics.com/imaging/cam ... meras/3398 AVT Prosilica GE-4900 35mm Color CCD Camera

or

http://www.edmundoptics.com/imaging/cam ... meras/3302 Prosilica GX6600 35mm CCD GigE Color Camera

[carllooper]

The cool thing about the multi-sampling approach you are using is that you can obtain the increased resolution of what a higher pixel dimension camera would otherwise give you, but without the cost of such cameras. There is some cut off that was calculated by some researchers in 2004

Under practical conditions the limit on SR enhancement was found to be about 1.6X, while under synthetic (or ideal) conditions the maximum limit was found to be 5.7X

So with a 1600 x 1200 camera one could induce, through multi-sampling, up to a effective signal rez of:

2560 x 1920 (practical) and
9120 x 6840 (theoretical maximum)

But this is purely in the spatial domain. Once you include temporal relationships (inter-frame correlations, similar to that used by some denoising strategies but without actually denoising*) you can start to push that spatial limit further out - and all without changing your acquisition camera - albeit at the expense of increased computation time of course.

Another example: if using the Retro8, with it's 720p capture camera, one could do some practical multi-sampling on that - which would bump the result up to at least 1152p.

C

* the visual result is still a decrease in noise, but without any increase in blurring.

[MattiasN]

ok

have you test to do multi capture and merge them to singel clip ?

i guess that Canon DSLR camera you have does have big sensor i mean it have good dynamic range so you do not need

[carllooper]

ok

have you test to do multi capture and merge them to singel clip ?

i guess that Canon DSLR camera you have does have big sensor i mean it have good dynamic range so you do not need

I'm getting 16 bit 5K images, so the DR is quite good - but it would be even better using the multi-sampling you've been doing. So I'm going to do that as well. Personally I don't think there is any limit to what is possible.

The work I've been doing has been concentrating on the inter-frame integration process - computing very dense optical flow fields - and merging information across time (and sorting what should merge and what shouldn't) - same sort of principal as multi-sampling a single frame but a lot more complicated by the movement (or differences) which occur from one frame to the next. Some changes are quantitative (differences in degree) and these are the useful ones. But others are qualitative (differences in kind) and if included in a merge they will muck up the results.

But if I also do what you are doing at the capture point, ie. multi-sampling the same frame - it will improve the original signal that is going into the inter-frame process. I've always been meaning to do multi-sampling but seeing the actual results you've achieved (rather than just mathematical proofs) it's given me the inspiration to actually make sure I put that in.

I'll have a three minute film by mid February that demonstrates some of the work I've been doing. Only some of it will be about faithful enhancement - whereas other parts will be more experimental - derived from mistakes I've made (if mistakes is the right word). Philosophically I'd argue anything one does in practice (or in reality) is not a mistake. Because reality, I'd argue, is not an error, ie. it's only in theory one can make mistakes.

C

[MattiasN]

16 bit 5K images that would be nice to capture

inter-frame integration process does sound complicated hmm is it like stacking photos like in macro shots i have seen when i have google about lenses

[carllooper]

16 bit 5K images that would be nice to capture

inter-frame integration process does sound complicated hmm is it like stacking photos like in macro shots i have seen when i have google about lenses

Imagine a movie in which each frame is of the same scene and nothing is moving. You could easily merge all the frames in the same way you are doing and get a higher definition (DR and/or rez). But any movement within the scene complicates it. So one has to determine for each pixel in one frame it's corresponding location in subsequent frames (and in previous frames) - and doing the merge along those movement lines. And making sure any incorrect correspondences don't get caught up in the integration (merge).

C

[MattiasN]

oh ok it was complicated

[JeremyC]

More great work Mattias.



ps. I've been using a Canon DSLR in my own work but am thinking of getting the 4K version of these (under $1K):

http://ww2.ptgrey.com/USB3/Flea3

At 4K it has a data rate of 21 fps. At 3.5 K (ie. transferring a sub-region of the sensor buffer) it'll run at 24 fps. My Canon, while 5K, runs at something like 1 fps when doing on the run RAW transfers.

Carl,

I would be interested in how well a CMOS sensor performs compared with a CCD sensor. However, wrt to the Flea 3 you linked to its pixel size is only 1.5 um and people more technically quick than me have pointed out that very small photo sites demand an extremely good lens otherwise having all those extra photo sites is useless and you might as well have something like a 1.5K sensor. I would love to get hold of a 4K sensor to build my own telecine but the only thing I've seen so far where the pixel sites are 5.5 um and larger and so correspondingly more expensive are the Kodak ones but they are CCD and whtehr the latest CMOS sensors can match them is another thing I need to investigate.

[carllooper]

Carl,

I would be interested in how well a CMOS sensor performs compared with a CCD sensor. However, wrt to the Flea 3 you linked to its pixel size is only 1.5 um and people more technically quick than me have pointed out that very small photo sites demand an extremely good lens otherwise having all those extra photo sites is useless and you might as well have something like a 1.5K sensor. I would love to get hold of a 4K sensor to build my own telecine but the only thing I've seen so far where the pixel sites are 5.5 um and larger and so correspondingly more expensive are the Kodak ones but they are CCD and whtehr the latest CMOS sensors can match them is another thing I need to investigate.

Yes, having a larger sensor pixel (ie. larger sensor array) will put less stress on the lens requirements (although you'd need a larger lens of course). But even where photosites are smaller than the purported resolution of a lens, the statistical sum of the sites, over the area designated as the rez limit, will give you a stronger signal in comparison to a single sensor pixel covering the same area. For the same lens , a 4K sensor pixel will still be giving you a 7X stronger signal per same area as the 1.5K sensor pixel. The strength isn't in terms of resolution per se, but in terms of bits per area, ie. in terms of dynamic range. And that can often be more important. To put it another way: a 4K signal downsampled to 1.5K will look much better than a signal originating at 1.5K - but it will also look sharper as well, despite both being 1.5K. There are legitimate reasons for this - something exploited in Super Resolution methods.

With multi-sampling you can increase the resolution of the lens mainly because the resolution limit of the lens will have been defined in terms of a single sample, but also that such tests are subjective. The reason subjectivity is (has been) required is because theoretical limits look (have looked) too crazy (higher than what seemed to be observable). But when applying machine vision methods the subjective tests will tend to have less meaning because the observer is not a human being - at least not at this point in the processing pipeline. And the multi-sampling method is at a tangent to the theoretical assumptions otherwise used in defining resolution limit metrics (such as MTF).

The multi-sampling (or SR) methods are also able to overcome (to varying degrees of course) the supposed limits of the original lens (and film stock) used during photography of the film. In the case of Super8 such lenses have not been the best - but using machine vision methods this limitation is not the limitation one might otherwise assume it to be. There are all sorts of innovative work being done in image reconstruction which put less stress on the hardware requirements (be they for new footage or archived footage). Of course one needs to be into that sort of thing to make any use of it. Alternatively one can spend tens (or hundreds) of thousands on a camera that may not need such methods. But for archive footage that isn't an option for the original taking camera. The digital processing methods become desireable as they can be applied not only to the lens and sensor adopted for transfer but the lens and film used at the time of original photography.

C

[JeremyC]

I see where you are going but I actually meant a 4K sensor with photosites of size 5.5 um or larger will be easier to find a lens for than a 4K sensor with photo sites of around 1.5 um when looking at exactly the same image. In that way the extra processing you mention is redundant and you can use the computer to carry out other image processing, such as noise reduction, for which there may be more pressing requirement. However, as far as my researches have shown 4K sensors with big pixels are expensive.