Conveying Shape:Lines

From CS294-10 Visualization Fa08

Jump to: navigation, search

Lecture on Nov 10, 2008




  • Automatic illustration of 3D geometric models: Lines. Dooley and Cohen. (acm)
  • Line Direction Matters. Girshick et al. (pdf)
  • Suggestive contours. DeCarlo et al. (html)

Optional Readings

  • Illustrating smooth surfaces. Hertzmann and Zorin. (html)
  • Automatic illustration of 3D geometric models: Surfaces. Dooley and Cohen. (pdf)
  • Speed of Perception as a Function of Mode of Representation. Ryan and Schwartz. (jstor)
  • Assessing the Effect of Non-Photerealistic Rendered images in CAD. Schumann et al. (html)

Seth Horrigan - Nov 10, 2008 02:17:17 pm

The Dooley and Cohen paper does outline a number of techniques that seem like common sense once presented but until then exist mainly in the ethereal realm of good drawing practice. Still something about the paper seems to suggest an overall lower quality than many of the other papers we have read. It is not merely due to idiosyncratic colloquial statements such as "Results from the system have been preliminary so far, but have been very exciting" or typos like "Difficulting relating to the inherent complexity of determining...". These elements perhaps contributed to the impression, as did the fact that the related works had only three items, but there seems to be some other elusive element. Nevertheless, it does provide useful information.

The paper on choosing line direction was thoroughly engaging. The paper opens with referencing psychology as the source of the interest in this topic: as such, I would like to see more information on how actual individual observers perceive the various objects. It seemed that the paper relied mainly on the authors' perceptions (with which I generally agree). I found it interesting that for the blob, I found the uniform vertical rendering easier to interpret than the principal or secondary direction vector, and roughly equivalent to the combined principal and secondary direction vector field rendering.

Kuang - Nov 10, 2008 03:00:54 pm

I'm really excited that I now know how to draw a doughnut, with the tapering lines for infinity and thickening lines for passing behind :).

The Dooley paper feels simple and obvious. In some ways, like good systems papers, the simplicity and duh-factor is why it's really good. I especially liked the figures in the paper; thought they are clear and well done. The Illustration Pipeline gives a useful breakdown of the illustration process wherein algorithms play parts.

One note, what happened to the rest of section 3?

Scott Murray - Nov 10, 2008 07:58:24 pm

I've enjoyed this whole discussion of lines because, well, they're just lines, right? And I never would have thought that so much thought and research has gone into the use of lines in visualization. Lines seem so un-sexy, yet they are incredibly practical for a wide range of applications, especially when dealing with display limitations (e.g. black and white only, low-resolution, static).

The Dooley and Cohen paper appeals to me for this very reason. Essentially, what they have done is to take a preexisting visual convention (using dashed lines and varied line weights to indicate shapes behind the frontmost shape) and translated it into algorithmic form. By formalizing what was already an informal yet familiar convention, the end result is very consistent and extremely intuitive for the end user. I wish that all line illustrations of 3D objects used this approach.

Ketrina Yim - Nov 10, 2008 07:27:53 pm

I found the concept of "suggestive contours" to be quite intriguing because it reminds me of the way I create line drawings (which also happen to be the most frequent type of artwork I produce). Unconsciously, line artists seem to practice the combining of contours with suggestive contours, probably because they know the limitations of using only contours. Drawings that employ contours alone tend to leave important details out, such as eyes, lips, or areas of negative curvature. One may not need the full outline of such details, but there has to be enough lines present to convey the information that these details have to offer, such as a character's expression. Much like the semi-schematic drawings, the combination of contours and suggestive contours is a compromise between using contours only and trying to capture as much of an object's curvature as possible (as in the Girshick paper).

Sarah Van Wart - Nov 12, 2008 12:34:55 pm

I also found the Dooley and Cohen article very interesting. On the complete opposite end of the spectrum from Ketrina, I don't believe that I've ever effectively drawn a 3D shape (well, maybe a box), and reading a critical analysis that "enlarg[ed] the vocabulary of lines" to explain how lines can be used to create effective 3D sketches was fascinating.

This hierarchy of line importance made sense -- first determine boundary lines and silhouettees, and then address curvature and determine an appropriate way to handle hidden lines. Similar to our "photo v. hand-drawn diagram" discussion in lecture, the article points out that using line drawings, as opposed to detailed, photo-quality renderings, is more useful (in certain contexts) to communicate the "full shape and intent" of a model. Indeed, sometimes less is more. I wonder if this additional knowledge will make me a better sketcher.

Michael So - Nov 15, 2008 08:59:26 pm

I found the study regarding the Perception of the 3D Configuration of Familiar Objects where they compare four types of images (photo, shaded, line, cartoon) to be interesting. The conclusion that the superiority of performance varies with the application is pretty vague. For imitating hand position, the photo was superior whereas for stating which switch was open, the cartoon won. I'm curious to hear a reason for those results. From the images of the switches in on and off positions, the photo image seemed the least clearest of the four images. I think it's due to lack of contrast in certain parts of the image. As for the hand images, the photo seemed to be the clearest, but in class everyone chose the line drawing and cartoon as the most likely to have the superior performance. I guess because people in the class felt a minimalist approach is usually the most effective. I'm also curious as to why the line drawing is consistently the worst in performance.

Witton Chou - Nov 16, 2008 02:15:54 am

Principal line directions is a good starting point in representing the shape of an object. While the lines provide the viewer with a sense of the contours of an object, it is not a representation that we are visually accustomed to. However, if combined with silhouette lines and shading, these visual queues can help enhance and clarify 3D volumes efficiently.

I really appreciated the suggestive contours piece. It is interesting how there are times when having too many contours will clutter the image whereas extensions to certain contours will enhance the perception of the surface by conveying smoothness. By exaggerating select contours, the geometry of an object can be better understood as opposed to using strict contours. It is fascinating to see how our mind processes these visual cues and understand which contours will enhance and which will merely clutter the visual representations of an object.

Matt Gedigian - Nov 16, 2008 08:24:36 pm

I wonder if the same semantic line styling that Dooley uses to create informative sketches from 3D models could be used in the opposite direction. Rather than having the computer communicate 3D shape to the viewer, could this be used to help an digital artist create 3D models from 2D sketches? Some of these visual styles are based on things that 2D illustrators already do, so perhaps there is nothing more that can be done. But I don't know whether there are currently tools that allow illustrators to specify the semantics of different lines (e.g. the meaning of a particular style of dashes line), which would be useful in trying to infer a 3D scene from a 2D projection.

Chris - Nov 16, 2008 07:18:43 pm

It would seem to me that there is a strong cognitive basis for the efficacy of line drawings in conveying information. In computer vision research, the image segmentation problem could be viewed, in some sense, as the problem of trying to create a line rendering from a real image. As an example, check out figure 15 (on page 16) of one example segmentation paper, Learning to Detect Natural Image Boundaries Using Local Brightness, Color, and Texture Cues. The output of the segmentation algorithm bears a strong resemblance to a line rendering of the scene.

This brings up a handful of research questions.

One question (which I would hope has been tested by now, but I can't find any references for it) is: Does visualization using a line rendering result in a different cognitive response than visualization using real images. In particular, were you to strap a bunch of humans into an fMRI and visualize their cognitive response to the two classes of images, would you find a noticeable difference?

The second question, assuming the answer to the first is "yes," would be: What is the nature of this difference? In particular, is it the case that using a line rendering allows the human visual system to "skip" doing extra processing to extract that data through segmentation? If this is the case, why is perception the line rendering better than the image rendering? Is it because the line rendering selects the relevant details rather than relying on the human visual system to decide on them (and possible make mistakes), or could it also be that there is less of a "cognitive load" due to being able to skip or ignore the extra processing to do segmentation? In other words, is it the case that the brain does some kind of load-balancing (an assertion I'm in no position to make), and that varying input can affect the effectiveness of later stages?

Maxwell Pretzlav - Nov 16, 2008 11:04:16 pm

Regarding "Line Direction Matters", like Seth, I found the uniform rendering of the abstract blob easier to interpret than either of the principal direction renderings. I also found the view in Figure 12 using both first and second principle direction vector fields to be clearer, despite the authors' claim that it should not be used and that "the inelegant crosses can become distracting and middle the flow of curvature." Possibly this would change if it were an animated view ... but for a still image I found these two renderings much clearer than the supposedly superior single principal direction vector fields.

I imagine this might be related to the fact that I was not initially familiar with the blob form at all -- the figures of forms I was more familiar with, such as pears, rabbits, and the horse, managed to very successfully convey the detailed shape of the 3D objects. I wonder if there is a difference in visual interpretation between seeing detail in overall shapes I'm familiar with compared with completely new unfamiliar shapes, and if different line algorithms may work better in one case than the other.

Calvin Ardi - Nov 17, 2008 02:08:34 am

Maxwell brings up a few points/comments I had while reading through Line Direction Matters.... I found interpreting the "blob" (figure 1) to be a lot easier in it's surface mesh form as opposed to the principal/secondary direction vector fields. However, I wasn't familiar with the blob; the vector field representations of the horse (fig 15) and bunny (fig 8) were certainly more recognizable (although the vector field of the bunny seemed too "busy"). The image of the pear (figure 14) loses some of the detail (the stem) that could be misinterpreted by those who may be unfamiliar with that fruit (a far stretch, probably, but something that should be taken in to consideration).

It was particularly interesting (and certainly relevant) that the importance of such methods were backed by psychological evidence (section 2). Does a system exist such that these images can be rendered on the fly to allow users to switch from images to these line drawings (or juxtapose the two together side by side)? This kind of ability (if I'm not mistaken) is implemented in 3d modeling programs or CAD tools, but what about one for medical imaging?

David Poll - Nov 17, 2008 06:51:31 am

I was particularly impressed by the notion of suggestive contours. The DeCarlo paper (and the corresponding website) describes these contours as "an extension of contours to account for 'nearby' viewpoints. I think this is incredibly apt, and I wonder why suggestive contours are so effective at rendering images using lines that we perceive to be more closely adherent and suggestive of the real image? The first thought that popped into my head here was that, in reality, we don't see from just one viewpoint -- we see from two (each eye). The brain then interpolates those images, giving us depth perception. I wonder if perhaps suggestive contours more closely align with the images our brains create of objects, and as a result they are especially effective at portraying a scene, especially when compared to other line drawings such as contours or silhouettes?

Granted, these other drawings show less data on a fundamental level, but I wonder if the threshold for interpeting these line drawings as real 3-dimensional shapes is reached once the amount of data exceeds the needs of some perceptual model in our brains related to the existence of multiple viewpoints? An interesting way to test this theory would be to compare the effectiveness of these line drawings for a person who is blind in one eye (and has been since birth) and one who has full vision.

Sorry for the side-track, I just got curious :)

Dmason - Nov 17, 2008 08:27:53 am

I wonder why no one has commented on the Hertzmann paper, which I found to be most engaging of the offering. For instance, I think everyone should be familiar with Figure 8 with different optimized hatches for Venus. They explain in clear language that most automatic hatching algorithms relate too complex of information. This comes down the the wildly sensitive principle curvatures they are usually based on. The paper's optimization routine weeds out these sensitivities, so that while its hatching isn't as accurate, it conveys larger-scale contours much better while relegating smaller-scale sensitivity to the intensity of the hatching.

Moreover, their discussion of Mach bands (explained here: points out a really obvious much much-overlooked point in illustration. Namely, that an illustration must capture not only the optical character of the object, but also the optical ILLUSIONS that we associate with the object. I find it fascinating to discover that my mental picture of something automatically incorporates these effects, and probably for good evolutionary reasons that I don't understand.

Razvan Carbunescu - Nov 17, 2008 02:03:33 pm

I am still mildly surprised at the fact that in the study of imitation of a particular position for the hand the line drawing faired so poorly amongst all options since for me at least it seemed to be the best depiction of the actual pose of the hand without encoding any other superfluous information but still containing a level of realism. I also have to agree with Ketrina that the suggestive contours of an object seem to be a very interesting concept and the program presented in class that automatically generates the suggestive contours seems like a very useful tool for rendering the general shape of an object without using too much information (I imagine all line drawings produced can be compressed significantly to reduce the space of saving a particular image)

HeatherDolan - Nov 17, 2008 02:56:52 pm

I found the Dooley and Cohen paper useful as well. The vocabulary of lines and the rules they have compiled is a helpful reference. I particularly liked the point about line endings. While it seemed obvious once I read it, I still found it helpful.

Matt, interesting question that you've asked regarding whether or not this would aid in the generation of 3D content from 2D data.

Jeff Bowman - Nov 17, 2008 12:42:51 pm

Like David, I too was most impressed by the idea of formalizing the suggestive contours. I would also argue that implicit suggestive contours is really responsible for the appeal of the surface mesh view of the blob. In fact, looking at the Girshick paper, I see the most (personally) attractive images as being images where the evenly-spaced contour lines indicate folds, and some element of faux shading from the foreshortened view of those surfaces. In fact, neither article talks about how the suggestive contours and contour lines is an analog for real-world shadows, which is what (it seems to me) really makes the difference.

I also wonder why we haven't seen a similar type of suggestive contour illustration in computer graphics. I see a lot of "toon-styled" computer animation on TV, and wonder if this is just something waiting to be implemented.

Finally, if you've stared too long at 2D shaded/contoured images for a while, try this.

Yuta Morimoto

"Line direction Matters" reminds me of some kind of ascii art. Surely, the paper approached in different ways from ascii art. Their start point is the flow of curvature over the surface and they try to transform 3D data to 2D ones by taking advantage of natural of curvature. I think their approach may be useful to emphasize a 3D model adding contour vectors generated itself. Because, additive line direction can emphasize 2D model. So when it applied 3D model additively, it might emphasize original 3D model by itself.

One thing surprised me in the lecture is about speed of perception on various representation of shape. As we see in the class, the best performance is drawn from photo or cartoon: thus realistic view or cognitive view. The performance is pretty much depending on the kind of environment or tasks such as case of surgeons shown in class. I think it is very straight forward and natural result. However, we had different opinions about speed of perception on hands in position in slide #14. We know our intuitions are different from the experimental or scientific result in some cases, but in many cases, our evaluations of visualization depend on very qualitative measures. So, the slides was good for reminding me of scientific approaches on visualization.

James Hamlin - Nov 17, 2008 04:35:52 pm

I'd like to briefly return to the question from lecture about what each of the photographic and illustrated visualizations of the pelvis convey. My problem was this: if any sort of operation, transformation, or whatever were permitted in the photographic case, then there's nothing to distinguish the possible end products except for the details of their production. We could produce a non-photorealistic image from the photograph and bring in domain-specific knowledge to do anything we can do in the illustrated case. Of course, if the question was 'what did that particular photograph and that particular illustration convey?', this is irrelevant.

On suggestive contours, I think they're effective because they stand in for the actual effects of parts of surfaces with normals nearly perpendicular to the view vector - compressed texture, view-dependent surface properties (e.g. specularity), and of course the gradients of direct lighting that occur across these areas of curvature.

Simon Tan - Nov 18, 2008 02:57:12 am

The paper formalizing suggestive contours for conveying shape has no discussion of the psychology behind the apparent need for suggestive contours like the paper on line directions, but I'm sure they could have shared similar literature as references. Both reflect aspects of the human visual system that help us determine textures of the things we see.

Artists seem to have a natural understanding of this, as both papers refer to their 'centuries' of techniques and innate ability to fill in these features without thinking. Ketrina confirmed this as she mentioned she does the same. I am not an artist, so although these papers may seem 'obvious' to some, they reveal aspects of artists' techniques that I never really noticed before. I wonder what difference there is in our brains that seem to separate the artists from the non-artists.

[add comment]
Personal tools