Vector Magic from Stanford (creates vector images from bitmaps)

[patrickl]

I tried some example from the galaga bezel, but there is one big problem with this tool for using it on artwork. It really does not deal well with big files (or big images). I tried a large chunk of the bezel and the result was seriously poor. When I give it a small piece of the bezel the result is a lot better (like above). Not sure what happens. Maybe they assume bigger images need less detail?

This really undermines the use of this for doing something like sideart. It would make it suitable only for doing logo's and characters.

[WunderCade]

This is a BETA version, so I wonder if they're gonna package it as download-able software. If they do, it will probably have the file size restriction removed.

[ChadTower]

Guessing from how they have it implemented now, I'm going to say it won't be downloadable.  It's a web service.

[WunderCade]

I went ahead and contacted them asking if they'd have someone click over here and answer some of our questions.....we'll see.

[patrickl]

This is a BETA version, so I wonder if they're gonna package it as download-able software. If they do, it will probably have the file size restriction removed.

It's not just the image size restriction. The file I uploaded was well within the limit. It's just that uploading a small piece of a larger image gives a different result as the same piece in the vecotrized result when you give it the full image. This means you need to cut your image in a lot of small pieces to get a satisfactory result.

[zorg]

I tried some example from the galaga bezel, but there is one big problem with this tool for using it on artwork. It really does not deal well with big files (or big images). I tried a large chunk of the bezel and the result was seriously poor. When I give it a small piece of the bezel the result is a lot better (like above). Not sure what happens. Maybe they assume bigger images need less detail?

they downsize large files, thus there is a loss in details.

[WunderCade]

So Patrick, wouldn't a full program be able to vector the image at 100% size? I mean, just because you could only see what's on screen, it should be able to vector the image at 100% size and then you could shrink/scale it down to see it all on your monitor.

The source image you input into the program would have to be at 100% actual size though. If you take a scaled down raster image and input that, then you're asking the program to vectorize an already degraded image. The key is to input a 1:1 image for best results, right? That's why I said it'd be nice to have a 6' x 3' scanner.

[patrickl]

It's obvious they downsize the uploaded files, but I'm not sure why. It depends if they downsample the files as part of the algorithm or if they are doing it to save diskspace/processing power/network bandwidth. If they just are trying to save resources indeed a downloadable version would not have that issue.

On the other hand, with other vectorizing software I have tried I first had to go through tons of settings to get any useful results. VectorMagic works with hardly any settings. So perhaps they downsample as part of the algorithm. It makes sense that they might be expecting larger images to be of higher resolution and thus assume they could be more loose with their vectorizing.

It's very likely that they downsample to save resources, but it might also be part of the algorithm.

[patrickl]

The source image you input into the program would have to be at 100% actual size though. If you take a scaled down raster image and input that, then you're asking the program to vectorize an already degraded image. The key is to input a 1:1 image for best results, right? That's why I said it'd be nice to have a 6' x 3' scanner.

Make sense yes. Usually the higher the resolution the better it will look.

With this type of software you never know though. Sometimes it actually performs better with slightly fuzzy images. Especially if you have flowing shapes (or circles) instead of sharp angled art. Still you'd probably be better of with slightly fuzzy high resolution images rather than low res slighly fuzzy images 

[ChadTower]

They only reason they'd limit the functionality to a certain size would be to save resources... or perhaps worry about abuse of copyrighted material and they think this will stop people from trying to profit from the app.

[WunderCade]

FWIW, James from Vector Magic team emailed me back after taking a look at the issues we've brought here on this thread. He asked me to post his response on their behalf so here it is:
--------------------------------------------------------------------------------------------------------------
Hi [WunderCade],

Thanks for your email, and for your interest in our project.  I took a
look at your forum discussion and I'd be happy to answer the questions
that I saw there.  If you don't mind, would you be willing to post
them on my behalf?

1) There was some question about the comparisons
(http://vectormagic.stanford.edu/vectorize/comparisons) and whether we
cherry picked them.  We selected images that we thought fairly
demonstrated the differences between our algorithm and the algorithms
available in the Adobe and Corel products.  In some cases all the
programs did essentially just fine and we avoided including those
images, since it is not very interesting to see essentially the same
result three times.  In no cases from the 124 user-provided images we
tested did we find that the Adobe or Corel results were actually
better.  An important thing to keep in mind here is that we did not
set out to solve all vectorization problems.  We are mostly interested
in re-vectorization, where the original was at some time, and in some
sense, vector art.  We have therefore restricted our sample set to
those examples.  As noted on the comparisons page, we compared our
results against those of Adobe and Corel on 124 user-provided images
and in not a single case did they produce results that both used fewer
shape primitives and had a lower re-rendering error (measured in root
mean squared error).

2) As several posters noted, the site did experience some difficulty
in meeting user demand.  This is because we were posted on Digg.com
and got flooded with Digg users.  The site is back up now and should
be working fine.

3) The file size and image size restrictions are because we are having
to do all the processing on our servers, and given the high level of
user demand we've been seeing, we can't afford to process jobs that
are too large, as it will tie up one of our processors for too long.
That restriction may be relaxed at some point.  How large are typical
scans that would be of interest to this community?

4)  A user named zorg posted an interesting comparison.  He seems to
have used different settings than I would have selected (indeed, if he
would provide a link to the result that he posted, I could perhaps
clarify which settings he might want to experiment with).  Here is a
link to the result that I got when I tried his source image:

http://vectormagic.stanford.edu/vctr/vctr_flex?g=37532&k=j481UGXJRae6nNOe&p=g

I've also attached the resulting EPS.  The results are not perfect, of
course, but I also don't think that a low-res low-quality JPEG with
sub-pixel-width outlines is typical input.  Indeed, most of our users
are providing much higher quality input and are getting much higher
quality output.  It is also important to select the options correctly.

If you would be so kind as to post the attached EPS, I think it would
help clarify the issue.  I would also be interested to see how well we
do on the same image, but at a slightly higher resolution and saved as
a PNG rather than a JPEG.

Regarding the broader issue of whether our tool is a substitute for
human vectorization, I would give the following remarks:  Human
vectorization will always remain essential for difficult images in
which the quality of the input is too low for a computer to do the
tracing, or images that were not originally vector art in spirit, and
require judgement and interpretation to produce a good result.  Our
tool is meant to help convert fairly clean, fairly high-quality
images, and to provide a starting point for marginal cases.

If you or any of the other posters have any further questions, please
feel free to email me at this address.  We are always trying to
improve the tool and welcome constructive feedback that will help us
do just that!

Very Best,
James

[ChadTower]

Excellent. 

[patrickl]

Cool that they respond.

We are mostly interested in re-vectorization, where the original was at some time, and in some sense, vector art.

That makes sense yes. Should be good for arcade artwork too right?

file size and image size restrictions ... we can't afford to process jobs that are too large

Is there information available on what kind of image size we should use for optimal results? I guess they cap the images at some fixed size. Which size is that?

For instance I uploaded a 969x828 image and it looks like it was downsampled to about a quarter of that resolution (when I zoom in on the bitmap image in the viewer). 23 pixels in the viewer relate to 40 pixels in my uploaded image wich would indicate that vectormagic is limited to a 0.25 Megapixel image (or around 557x476 in my example)

Or is only the viewer limited and is the bitmap actually used at full resolution?

Right now I don't know what to do to get optimal results.

BTW the Galaga bezel is 23"x 21" scanned at 300DPI and is about 44 Megapixel in total. Sideart scans will probably get even be bigger than that.

:edit:

Perhaps they could check some samples from localarcade? For instance the Galaga bezel is http://www.localarcade.com/arcade_art/r857.search.htm

[shardian]

Well now that I have messed with it a while, it is a very neat program, but also kind of frustrating. Still, it has now been extremely helpful in getting something I needed badly.

My wife custom made a Marshall cheerleader outfit for our little girl. I needed a logo graphic for an iron on, but have never had any luck finding a high enough resolution graphic to keep it from looking like crap. Using this software, I have been able to make a quite good, though not totally accurate logo.




As you can see, there are glaringly obvious errors in the horns and eyes. I did my best, but find the editing function very lacking. I made the two eyes perfectly identical in the edit mode, but the program still interpreted them differently. Maybe I am missing something, but there is no way to edit any layer but the top most. Simply painting over does not help when it comes to re vectorizing. I did have a little bit better version, but had unacceptable problems getting rid of some clutter at the bottom.

So my final evaluation of this program is it can be very helpful and quick for small projects that simply do not have a decent resolution image available. Still, it is not on the same level as a complete vectorizing package. However, put into the perspective of cost, this software kicks ASS.

[ChadTower]

When you consider that took probably 5 minutes of real time effort on your part it is a ---smurfing--- fantastic result.

[shardian]

When you consider that took probably 5 minutes of real time effort on your part it is a ---smurfing--- fantastic result.

Well in reality I tried with varying success with about 6 different low res images. In all, I'd say I spent a little over an hour total.

[shardian]

Oh, and I forgot to mention it does not handle text all that well. I mean it did okay. I could read it and all, but it just looked weird and unnatural. It would be neat if you could highlight a part of the picture and tell it "there is text here. Analyze as such".

[ChadTower]

I bet if you eliminated the artifacting in the source it would... maybe take that section of the image and reduce it to two colors.

[patrickl]

I also noticed it doesn't work very well on text. Even on a 300DPI scan.

[ChadTower]

Text, in a lossy format, has too many jaggies for a vectorizer to be perfect on... you'd have to clean up those jaggies manually before putting it in.  The question is would it be easier to clean up the jaggies in the source or to fix the vectorized output.

[patrickl]

The problem is more that text is really sensitive. You can always tell if there is something wrong. When you trace some irregular shape, people will only see that you got it wrong if they compare it to the original artwork. When you look at text you can simply tell that it's wrong by looking at it.

It's seriously difficult (timeconsuming) to get it right when you trace it by hand too. I've tried to trace text a few times, but always just looked up the font and used that.

[shardian]

Text, in a lossy format, has too many jaggies for a vectorizer to be perfect on... you'd have to clean up those jaggies manually before putting it in.  The question is would it be easier to clean up the jaggies in the source or to fix the vectorized output.

I did correct the text in their editor. But like I said before, their editor leaves much to be desired.

[ChadTower]

Illustrator isn't hard to get... or checkout Inkscape, it's not bad, and it's free.

[HaRuMaN]

They output .svg files, right?  Just grab SodiPodi...  its a free vector program, and its pretty good.  I used it to create my two entries to the arcade art library.

[WunderCade]

Does anyone know of a way or a program that will convert an .ai file into and .svg file?

[XyloSesame]

Does anyone know of a way or a program that will convert an .ai file into and .svg file?

Inkscape will open an Illustrator ai - you can then save the file as an svg.

Illustrator will, of course, as well.