6 ms·

Photo-Sketching: Inferring Contour Drawings from Images

54 points by pplonski86 8y ago

ximeng 8y ago
Quality doesn't seem much better than the comparison algorithms in the paper.
- jnbiche 8y ago
  Which images are you looking at? The ones I've seen appear dramatically different and better to me(but I'm also a hobbyist artist and sketcher).
  - topspin 8y ago
    Likewise. The the new technique produces dramatically more distinct edges, closing gaps where HED et al. taper off and finding meaningful interior edges that others omit without also pulling in noise edges.
  - ximeng 8y ago
    E.g. on page 7 of the paper where the head is cut off in the "ours" picture and less so in the "HED" picture, or the helicopter is similar or worse in quality to the "gPb" picture.
    - jnbiche 8y ago
      Sure, it's not 100%. They included those images to illustrate some challenges that they ran into (negative space and low value boundaries, not surprising given their relatively small data set). But the ones that they open with are dramatically better than HED.
      - ximeng 8y ago
        Hmm, I'm getting downvoted, but even the opening ones look arguable to me, e.g. in the hotel room on the first page the lamp shade and cushion near the sofa don't look great. Some of the HED artifacts could be improved with some post processing I imagine (as they do later). The dogs are significantly better on page 6. Looks to me like there are some things they are doing better, but still a way to go. Difficult to tell with the small number of sample images.
        jnbiche 8y ago
        I didn't downvote, but I still see a significant difference between HED and the author's algorithm. But you're correct, it is difficult to make a confident judgment given the small number of sample images.
jnbiche 8y ago
Check out the pdf, pretty impressive results: https://www.ri.cmu.edu/wp-content/uploads/2019/01/Li-Mengtian-WACV-2019-Photo-Sketching.pdf https://www.ri.cmu.edu/wp-content/uploads/2019/01/Li-Mengtia... BTW, if the authors find these comments, they should definitely advertise the new game they create to collect more data (once it's released) to art-centered communities, particularly comics. Many people who do line art practice tracing frequently to improve their technique.
JeremyBanks 8y ago
It's 2019. Can we please stop using the Lenna Playboy centerfold picture in every other graphics paper? Perhaps in the 70s there was actually a shortage of useful reference images and it had some value -- I wouldn't know -- but at this point it's just creepy.
- kgraves 8y ago
  Agreed, it's 2019 I too would rather have the computer vision community use a different image instead Lenna, perhaps the proposed Fabio Lanzoni[1] or the Kodak image suite[2]. [1] https://www.cmc.edu/news/every-picture-tells-a-story https://www.cmc.edu/news/every-picture-tells-a-story [2] http://r0k.us/graphics/kodak/ http://r0k.us/graphics/kodak/
  - jessaustin 8y ago
    It's not clear why the objections lodged against the original image could not also be lodged against Kodak's "kodim04" image. Both are images generated in the 1970s that portray smiling attractive young women of Scandinavian descent with bare shoulders. [EDIT:] take a look and tell me I'm wrong: http://r0k.us/graphics/kodak/kodim04.html http://r0k.us/graphics/kodak/kodim04.html
    - p1necone 8y ago
      Because the original image was cropped from pornography.
  - cookingrobot 8y ago
    I like the idea of changing it, but neither of those options seem good. Fabio is just a response to Lenna - it seems like we can do better. And those Kodak images are really surprisingly bad. Consistently drab colors, blown highlights, and really badly composed pictures of dated subjects. Is there something I'm missing that makes those images important?
    - jessaustin 8y ago
s1mon 8y ago
Hmmm. The examples are underwhelming compared to commercial examples of these types of filters. iOS 12 added filters which are only available in Messages or the Clips app (not the camera app) that produce amazing results in realtime. Try the "ink" version. https://www.cultofmac.com/583400/how-to-use-ios-12-comic-book-photo-filter/ https://www.cultofmac.com/583400/how-to-use-ios-12-comic-boo... For even more extreme versions, there are tons of filters, plugins etc for Photoshop and other paint programs. Here's one from 2013: http://www.tipsquirrel.com/the-graphic-novel-filter-photoshop-elements-diamond-in-the-rough/ http://www.tipsquirrel.com/the-graphic-novel-filter-photosho... Maybe I'm missing something, but why is it that academics manage to publish stuff like this that is so far from the state of the art in the commercial world?
- jnbiche 8y ago
  > Maybe I'm missing something This may be the case. The filters you cited are not countour filters. Having the option of pixel matching entire regions by color makes any image look better, regardless of how many or how few contours you match.
- SiempreViernes 8y ago
  Their code make actual outlines, it doesn’t merely detect edges and applies a transform to those pixels. Basically they can make inks automaticelly, if poorly.
- csmeder 8y ago
  FYI it seems like this filter relies on Apples’s latest depth detection hardware not just an image. I have a pre X phone and that option doesn’t come up for me.

54 points by pplonski86 8y ago

ximeng 8y ago

Quality doesn't seem much better than the comparison algorithms in the paper.

jnbiche 8y ago
Which images are you looking at? The ones I've seen appear dramatically different and better to me(but I'm also a hobbyist artist and sketcher).
- topspin 8y ago
  Likewise. The the new technique produces dramatically more distinct edges, closing gaps where HED et al. taper off and finding meaningful interior edges that others omit without also pulling in noise edges.
- ximeng 8y ago
  E.g. on page 7 of the paper where the head is cut off in the "ours" picture and less so in the "HED" picture, or the helicopter is similar or worse in quality to the "gPb" picture.
  - jnbiche 8y ago
    Sure, it's not 100%. They included those images to illustrate some challenges that they ran into (negative space and low value boundaries, not surprising given their relatively small data set). But the ones that they open with are dramatically better than HED.
    - ximeng 8y ago
      Hmm, I'm getting downvoted, but even the opening ones look arguable to me, e.g. in the hotel room on the first page the lamp shade and cushion near the sofa don't look great. Some of the HED artifacts could be improved with some post processing I imagine (as they do later). The dogs are significantly better on page 6. Looks to me like there are some things they are doing better, but still a way to go. Difficult to tell with the small number of sample images.
      - jnbiche 8y ago
        I didn't downvote, but I still see a significant difference between HED and the author's algorithm. But you're correct, it is difficult to make a confident judgment given the small number of sample images.

jnbiche 8y ago

Check out the pdf, pretty impressive results: https://www.ri.cmu.edu/wp-content/uploads/2019/01/Li-Mengtian-WACV-2019-Photo-Sketching.pdf https://www.ri.cmu.edu/wp-content/uploads/2019/01/Li-Mengtia... BTW, if the authors find these comments, they should definitely advertise the new game they create to collect more data (once it's released) to art-centered communities, particularly comics. Many people who do line art practice tracing frequently to improve their technique.

JeremyBanks 8y ago

It's 2019. Can we please stop using the Lenna Playboy centerfold picture in every other graphics paper? Perhaps in the 70s there was actually a shortage of useful reference images and it had some value -- I wouldn't know -- but at this point it's just creepy.

kgraves 8y ago
Agreed, it's 2019 I too would rather have the computer vision community use a different image instead Lenna, perhaps the proposed Fabio Lanzoni[1] or the Kodak image suite[2]. [1] https://www.cmc.edu/news/every-picture-tells-a-story https://www.cmc.edu/news/every-picture-tells-a-story [2] http://r0k.us/graphics/kodak/ http://r0k.us/graphics/kodak/
- jessaustin 8y ago
  It's not clear why the objections lodged against the original image could not also be lodged against Kodak's "kodim04" image. Both are images generated in the 1970s that portray smiling attractive young women of Scandinavian descent with bare shoulders. [EDIT:] take a look and tell me I'm wrong: http://r0k.us/graphics/kodak/kodim04.html http://r0k.us/graphics/kodak/kodim04.html
  - p1necone 8y ago
    Because the original image was cropped from pornography.
- cookingrobot 8y ago
  I like the idea of changing it, but neither of those options seem good. Fabio is just a response to Lenna - it seems like we can do better. And those Kodak images are really surprisingly bad. Consistently drab colors, blown highlights, and really badly composed pictures of dated subjects. Is there something I'm missing that makes those images important?
  - jessaustin 8y ago

s1mon 8y ago

Hmmm. The examples are underwhelming compared to commercial examples of these types of filters. iOS 12 added filters which are only available in Messages or the Clips app (not the camera app) that produce amazing results in realtime. Try the "ink" version. https://www.cultofmac.com/583400/how-to-use-ios-12-comic-book-photo-filter/ https://www.cultofmac.com/583400/how-to-use-ios-12-comic-boo... For even more extreme versions, there are tons of filters, plugins etc for Photoshop and other paint programs. Here's one from 2013: http://www.tipsquirrel.com/the-graphic-novel-filter-photoshop-elements-diamond-in-the-rough/ http://www.tipsquirrel.com/the-graphic-novel-filter-photosho... Maybe I'm missing something, but why is it that academics manage to publish stuff like this that is so far from the state of the art in the commercial world?

jnbiche 8y ago
> Maybe I'm missing something This may be the case. The filters you cited are not countour filters. Having the option of pixel matching entire regions by color makes any image look better, regardless of how many or how few contours you match.
SiempreViernes 8y ago
Their code make actual outlines, it doesn’t merely detect edges and applies a transform to those pixels. Basically they can make inks automaticelly, if poorly.
csmeder 8y ago
FYI it seems like this filter relies on Apples’s latest depth detection hardware not just an image. I have a pre X phone and that option doesn’t come up for me.