Conveying Shape: Lines
From CS294-10 Visualization Sp11
Lecture on Apr 20, 2011
- 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)
Saung Li - Apr 20, 2011 06:08:58 pm
The Girshick et al. paper provides a compelling argument for the use of principle directions to better convey the shape of an object. The direction of lines does indeed have an affect on our perception of shape: figures 2-6 provide illustrative examples of why using principle direction vector fields are more effective that other vector fields such as random and uniform ones. Adding silhouette lines further improves the quality of the image, as shown in figure 15, and it would be interest to see shading and variable length lines to improve the image even more. For some of these images, I don't see what we gain from representing objects using vector fields. For example, in the horse figure I feel like I can better understand the shape of the horse from a visualization such as a cartoon version of it that's colored in with its fur flowing in the direction of the vector field. From this I can still understand the shape of the horse and it will look better aesthetically and more realistic.
Michael Hsueh - Apr 22, 2011 12:44:30 am
I agree with Saung's assessment about the effectiveness of lines of principal direction. It was not initially obvious that they would be better than random vector fields, as there is a tendency to suspect that we might omit important information by focusing entirely on principal directions. Random vector fields almost seem as if they could provide more "coverage" of the general shape of objects due to their random nature. Judging from the results shown in the paper, this is simply not true, as more information comes from the varying density of lines shown (what amounts to some sort of pseudo-shading) than the shapes of the random lines themselves. Anyway, one of the key features of using principal directions is their geometric invariance, allowing animation of drawings without the distraction of view dependent details such as shading. By introducing view-dependent variables such as variable lighting or shading that are either guided or combined with principal direction vector lines, we can get a very nice representation of almost any shape. These additional details might help in some ways with dealing with difficult regions of opposing force or undefined principal direction.
Matthew Can - Apr 22, 2011 03:16:39 am
The problem of deciding which lines best convey shape is an interesting one. The work on suggestive contours is a solid contribution toward addressing this. But what I really would have liked to see is a formal user evaluation that compares this to other methods such as lines of principal curvature. The paper's brief, informal discussion doesn't satisfy me. But perhaps the reason they left out a formal evaluation has more to do with the expectations for graphics papers (as opposed to visualization or HCI) than anything else. In any case, one thing that caught my attention is their algorithm for computing suggestive contours from images. Unfortunately, they only showed results on images that were rendered from triangular mesh objects. I wonder if the algorithm would work on images captured with a digital camera.
Manas Mittal - Apr 22, 2011 11:15:39 am
All in all, very interesting papers.
I noticed that these mechanisms rely on a 3D model to be provided, of which, they'll do things are contour detection etc. This made me wonder aloud about 2 questions:
First, it is better to provide a picture as compared to a sketch for a 3D object? I know there was some discussion in class about it (and I remember the image of the animal pelvis), but my question is for familiar objects. I am inclined to think about people who don't have binocularity in their eyes (i.e., they are unable to resolve the 2 images our eyes sees and render it into a 3D image). We should see if they have different perceptive mechanisms, i.e., humans adapt their visual mechanisms.
Second, we saw in a talk this past week at the BiD lab how designers can express 3D models using 2D Cad tools. There was a cool idea there of 'sucking the air out' of a mesh to get a 3D image. I am wondering if we can use some of our work here to not just visualize but to use it as a way to render. I can think of user study scenarios....
Sally Ahn - Apr 22, 2011 08:38:49 pm
Dooley and Cohen's paper discusses many attributes of illustrated lines, such as thickness and end conditions. I liked that they recognized these attributes because I agree that they impart important meanings for drawn lines. However, I was disappointed with their results because they did not demonstrate many of these attributes (varying thickness, tapered ends, etc.), but rather only demonstrated hidden line depiction. Moreover, they categorize lines into four groups: boundary lines, silhouette lines, discontinuity lines (folds), and contour/isoparametric lines. I wish they discussed this in more detail, because I am not sure what attributes separate the lines in each group; they don't seem mutually exclusive. The other two papers seem to focus only on the last group: contour/isoparametric lines. I would like to see more work on automatic selection of the line attributes (such as varying thickness and tapered ends) mentioned in the early Dooley and Cohen paper (their system requires some user input).
Siamak Faridani - Apr 24, 2011 10:27:34 pm
I liked the fact that almost all of the readings for this class had a computer graphics flavor to them. Hertzmann's paper was very interesting. I wish they had explained their procedure a little better though and I really wanted to see more examples of their results.
In Ryan and Schwartz's paper one thing that was missing (in my opinion) was a discussion on average reaction time for different people. I mean we know that people have different reaction time and for example computer gamers have a very fast reaction time. So the question is how big are these differences that they present compared to the reaction time for each person? Their statistical method is not really reliable either I think they should have at least used an ANOVA test to make sure that there is a difference between these models. They only provide the mean and there is no discussion on variance of these times. I refuse to accept that their reasoning is reliable as there is no evidence that what they really claim is a valid difference.
Karl He - Apr 25, 2011 01:52:35 pm
Line-techniques for conveying shape are interesting in that they use something that wasn't in the original image. Showing suggesting contours and principal lines add something that technically wasn't in the original 3D shape but help people understand the nature of the 3D shape. It is similar to the apparent-area technique used for displaying area described earlier in the class, however in this case it is not misrepresenting any quantitative data.
Michael Cohen - Apr 25, 2011 02:28:24 pm
I thought that my main objection to the Ryan & Schwartz paper would be that the results don't seem significant enough, but looking at the paper it appears that they were reasonably careful with their statistical tests (although they don't seem to report their n, so it would be hard to check their work). However, I notice that their procedure is a little contrived -- rather than presenting a single image until the correct answer is given, they present each image in very brief (sub-second) increments, asking the subject for the answer after each increment. If the subject does not know, or guesses incorrectly, they are shown another increment, and so on until they get the right answer. Thus, the results apply to real-world perceptual efficiency only if you assume that the brain processes ten distinct 0.01s presentations in the same fashion that it would process one 0.1s presentation. It may be the case that cartoon images are more suited to perception in these very quick "slices" because they are simpler and the viewer can perceive more in one "blink" -- but if given an uninterrupted viewing where the visual system can build up a continuous picture over the full viewing period, other methods may perform just as well.
Ryan & Schwartz acknowledge this problem and state that a better design would be to show each subject each image once for a random amount of time, but that this would require a prohibitively large number of subjects. However, another approach that could be run with a small subject pool would be to let the subject hit a buzzer when they believe they know the answer, and then exclude trials where the response was incorrect. This would be testing something slightly different (when the subject is confident she can give the right answer, as opposed to when she'd be able to give the right answer if pressed, whether she's confident or not) but I would argue that the confidence metric is actually more relevant in the real world. Except in the most dire emergency, a person will continue to look at an image until they're confident they understand it, even if they may have unconsciously reached the correct understanding slightly before that.
Brandon Liu - Apr 25, 2011 06:34:13 pm
The most revealing part of the Ryan & Schwartz paper was how the cartoon result was the fastest to interpret. What would the results have been with a different cartoon drawing? The authors described their technique for creating the cartoon as distorting the shape while preserving the silhouette. However, the shape of the thumb is more symbolic and like a swoop, while the fingers have had their segments reduced from 3 to 2. If the hands were further simplified, I hypothesize that the reaction times would have been slower. For example, a stick figure hand would be hard to interpret in this context. Thus, the specific choices made in the cartoon are there to communicate the spatial relationships faster.