Conveying Shape: Lines
From CS 294-10 Visualization Sp10
Lecture on Apr 12, 2010
- Automatic illustration of 3D geometric models: Lines. Dooley and Cohen. (acm)
- Line Direction Matters. Girshick et al. (pdf)
- Suggestive contours. DeCarlo et al. (html)
- 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)
Zev Winkelman - Apr 02, 2010 04:07:30 pm
This material seemed to be primarily concerned with representation of concrete objects that occur in 3D space. Though I found the work to be very engaging, I am more interested in representations of abstract data. Some of the findings could be applied to 3D models of abstract data, but this did not seem to be the focus.
Ryan and Schwartz:
Interesting methodology. Don't know that I would have guessed that the cartoons would be most quickly perceived, but after looking at the example with the switches, it seemed as though perhaps the cartoons used techniques that made the relevant details stand out - such as coloring the handles solid black - that didn't occur naturally in the image.
Dooley and Cohen:
Introduction of the idea that structure can expand detail in images in a way that photo realism can not.
Like the idea of varying line style from beginning to end, don't like the idea of having to specify all of that. Good thing these guys are working on the algos to alleviate some of that work. I like the final product (262)
Girshick et al.
More on NPR as perceptually efficient. Had never even thought about it. Guess I was stuck in a PR paradigm world, but I can see how some of these techniques are a definite improvement. Although figures 2-5 made me think of some kind of warped rorschach test. All jokes aside, the additional information communicated with the lines in figure 15, as opposed to just the silhouette, is substantial.
DeCarlo et al.
Here: "Lines are the scaffold of non-photorealistic rendering..." (1.1) Trying to remember where else ...is the scaffold... was in our readings but blanking. Suggestive contours seem like something I used to put in my drawings as a child (probably incorrectly) to accentuate a tree branch or a mountain (like figure 4a). Like Girshick's horse, figures 10,11,12 and 13 show a clear improvement over contours alone.
Hertzmann and Zorin
Lots in here, but same basic argument that NPR techniques (hatching in this case) can create more perceptually efficient images ?
Schumann et al.
An important distinction is drawn regarding the effects of NPR techniques - good for some things (suggestive diagrams leaving something open for interpretation), not good for others (final diagrams). I'm still trying to understand the design behind figure 7.
Jon Barron - Apr 12, 2010 12:40:01 am
Dooley: This paper lost a lot of oomph when I realized it wasn't about automatic contour generation. As it is, I'm not sure what the contribution of this paper is, besides assigning names to the attributes that contours have. Defining the edges of line segments in image-space seems like strictly worse than defining them in 3D, and producing "closed" line segment ends by proper Z-buffering.
Girshick: I like this paper. Principal directions definitely seem like a powerful tool for graphics and visualization, though I'm not quite sold on the line-rendering idea. It seems like a lot of what you get when rendering these lines is the same as what you get by simply rendering the mesh, lit from the direction of the camera (surfaces oriented perpendicular to the camera are heavily shaded, surfaces pointed at the camera are less shaded), though certainly more information is in the line drawings, if for no other reason than each polygon contributes two numbers ( x and y ) instead of one (shading). A comparison against renderings of the mesh would be nice.
This idea seems most compelling as a technique for deciding how textures (wood grain, for example) should be mapped onto unusual surfaces. This is effectively what they do in Figure 14.
DeCarlo: The crux of the idea seems to be this line: "Equivalently, they are those locations at which the surface is almost in contour from the original viewpoint —-- locations at which the dot product in is a positive local minimum rather than a zero." I like it, it's elegant and simple. Reminds me of the opposite of the Canny edge detector.
Paul Ivanov - Apr 12, 2010 04:48:17 pm
Dooley and Cohen: This paper is a good first step in addressing the valley between the the two delta peaks of either completely solidly rendered or completely wireframe options, which are so prevalent in interactive 3d modeling: it's hard to believe that it took until 1990 for such a contribution to be made. The main insight is that not all lines are equal - importance and meaning of lines differs depending on the user, applications, or task. So it is natural to allow the user to specify and attach the meaning/importance. Having a common vocabulary for the different kinds of lines is important, because it allows for a richer specification of line drawings, but this paper seemed to concentrate only on the depth-ordering aspect of discontinuity lines
It seems like this small-multiples matrix of the numerous possible intersections between objects (Fig 2), while great for interactive usage context from the user's perspective, can also be used in effectively communicating the user's specification for future viewers (as a kind of legend when the graphic is used in a publication, for example).
I also liked the Tron-like color plates :) .
Girshick et al: The case for using principle directions is argued well and convincing in writing. The close-ups in Figures 6b-d, in particular, demostrate the benefit of using principle directions, though the straw men or random and uniform vector field seem pretty dubious, and a surface-normal vector field comparison would have been nice to have for comparison. But of the specific line drawing image examples, only figure 10 looks reasonable. Why add noise to the stroke tracing process? (Figures 11d, 14) It looks really bad! If anything, because we're usually dealing with limited precision floating point numbers, it might be worth-while to *smooth* the curves, instead of introduce noise.
Stephen Chu - Apr 19, 2010 01:55:38 am
Last semester, I was introduced to the Waltz Algorithm, which is used to interpret line drawings of solid polyhedra. It classifies each line as a boundary line, interior convex edge, or interior concave edge. This algorithm is modeled as a CSP problem and makes assumptions that simplify the problem (objects have no shadows or cracks, 3-faced vertices, and junctions don't change with small movements of the eye). This was a good introduction to the power of lines in conveying dimension and shape, but I was left wondering how one would interpret line drawings of an object outside of the polyhedra set. The lectures and readings this past week have tackled much more complex problems that show the variety and importance of lines in our perception of 3d objects. I especially enjoyed the Suggestive Contours papers and Line Direction Matters.
Arpad Kovacs - Apr 19, 2010 02:55:24 pm
Dooley and Cohen: Many of the statements/findings of this paper are very simple, so it seems that the paper's main contributions are defining a concrete vocabulary and some rules of thumb for how to handle lines effectively. I was initially confused by the wording of the formula on page 80 (what is yon-hither?), until I realized that this is just a simple linear interpolation of line width between the min line width at the far plane, and max line width at the near plane. Likewise, they vaguely suggest a similar interpolation of the dash lengths ("dashes should be shorter as the line moves away from the eye, but not too short") although I would have appreciated more concrete, well-defined instructions/formulas. Nevertheless, I really liked the small-multiples matrix in figure 2, which very cleanly and efficiently encodes importance, line type, and hiddenness using only line thickness and dash lengths (line style).
Girschik et al: This paper discusses how lines derived from vector fields can reveal shape in nonphotorealistic rendering. Its main contribution seems to be that vector fields consisting of the first and second principle directions are more effective at conveying the local surface orientation, and thus the shape of an object, than random or uniform vector fields. It also shows how tracing a carefully selected set of streamline strokes through the vector fields, combined with hidden line removal and silhouetting can reveal 3D volume data.
Although the authors discourage the simultaneous use of first and second principal direction fields as shown in the "crosshair" approach of figure 12, I think that this crosshatching approach reveals the shape of the object better than figures 2 through 5. Since a crosshair at a given point will be locally tangent to the plane, the viewer can imply the normal vector at that point, which is much harder to do with a simple vector representation. If taken to the limit, where each crosshair's lines are extended to a neighboring point, then this approach approximates a very dense, 1-sided wireframe rendering, which seems to be quite popular in the CAD community for visualizing the underlying 3D structure of an object.
Prahalika Reddy - May 12, 2010 08:00:26 am
The issue of which is more effective, photographs or drawings, is a very interesting one. I feel both sides have good points. With a photograph, I feel it's sometimes easier to show the exact texture and color and overall detail of the object. However, with drawings, the author can highlight whichever characteristics are most important. Obviously, this distorts the actual image, but sometimes it's desirable to emphasize a certain portion of the image and therefore, a good idea to use drawings.
The concept of line weight is also an interesting one. It's interesting to see that mere lines can make a huge difference in the image itself, almost as much as color and shading. Proper line weights change the way an entire image is viewed; that's pretty powerful.
Shimul Sachdeva - May 13, 2010 10:41:08 am
Suggestive contours. DeCarlo: While suggestive contours are usually a great feature to have, especially in a hand-drawn image, an overdose kills the purpose and in fact loses its impact as being "informational".
The Klein bottle example showed in class reminds me of Professor Sequin's work on knots and the mathematical models that go behind these curvatures.
Dooley: Illustration rules enumerated in this paper are comprehensive and quantify the rules by giving them an importance value. The mantra of "enhance important features and deemphasize unimportant details" is useful and apt.