| Who | When |
Messages | |
|
|
|
| travestia
|
18
|
 |
|
07-24-2008 03:45 AM ET (US)
|
|
|
| |
Messages 17-16 deleted by topic administrator between 07-25-2008 02:06 AM and 06-25-2008 02:26 AM |
sunglow
|
15
|
|
|
06-13-2008 04:23 AM ET (US)
|
|
|
| sevgher
|
14
|
|
|
06-12-2008 03:40 AM ET (US)
|
|
|
|
|
13
|
|
|
05-16-2008 04:14 AM ET (US)
|
|
Deleted by topic administrator 05-16-2008 08:08 AM
|
| Nicole
|
12
|
|
|
07-21-2006 04:52 PM ET (US)
|
|
Looking for information and found it at this great discussion. There's also some on coreg shortage webpage devoted to coreg shortage. metformin and pregnancy webpage devoted to metformin and pregnancy. |
| Shinko Cheng
|
11
|
|
|
11-06-2003 07:16 PM ET (US)
|
|
Edited by author 11-06-2003 07:23 PM
Regarding Neil's question 2 about why the background shape layers sum to 1, lumping the background together into one pdf is a functional simplification. Doing so allows the front elements of the "background" to be considered the same class with a common motion as predicted by the motion model, so the algorithm will not need to assign the predicted "background" motion to each occluding "background" layer. One of the primary assumptions is that, that the "background" has a common motion model. I can imagine that we model the layer visibility probability for all the layers just as was done for each foreground layer, and then assume that every other layer is the background layer or some combination of layers be assigned the background, but that would mean an additional step is required to decide whether or not each layer contain the common "background" motion and classify each as such.
I like the idea of layers for segmentation. Specifically, I like the idea of using a deformable appearance model to distinguish between multiple objects in images. I can imagine that there would be a tradeoff between finding a model to describe appearance deformations to keep objects a single layer and allowing objects to be divided into different layers based on objects lacking fit with that particular apperance model.
|
| Mike McCracken
|
10
|
|
|
11-06-2003 12:33 PM ET (US)
|
|
I was curious about their use of 2d motion models. They say that background motion is described with a 2d affine model, and foreground layers undergo 2d rigid motion. This sounds like there would be some types of motion that would confuse the algorithm, including some camera motions - am I right, or is 2d enough? I just wasn't sure if they made that choice because it's enough or because it was some kind of tradeoff...
|
| Neil Alldrin
|
9
|
|
|
11-06-2003 04:31 AM ET (US)
|
|
A few questions.
1) They refer a couple times to the "direct method"; what is the direct method?
2) Why do they require the the shape maps for background layers to sum to 1? It seems like they could have removed this restriction with no harm done.
3) How well could layering methods be adopted to non-stationary cameras, where parallax would create non-uniform motion among all the background layers?
These questions are probably pretty stupid as I didn't read the paper as closely as I would have liked (starting before midnight would help :)).
|
| Meifang Huang
|
8
|
|
|
11-06-2003 02:23 AM ET (US)
|
|
The effect of shadows in Figure 4 is a very interesting problem. Normally we will consider the shadows as background layers,but when the object is moving under these shadows, the shadow would not be in the back of the object, instead, it will become a occluding background layer before the moving object and not completely occlud it. In this situation, the background layer model did help to solve this problem, because it can dynamically change the layer order, and assign a background layer along with a foreground layer to this semi-occluded object.
|
| Meifang Huang
|
7
|
|
|
11-06-2003 02:00 AM ET (US)
|
|
This paper is based on the authors' previous work of dynamic layer model, which extends the estimation of layer model to incremental estimation formulation. The key idea of this paper is that they use multiple background and foreground layers to segment the scene for tracking purpose. The multiple background layers would help to solve the complex background occluding problems and also do well for the object tracking.
|
| Piotr
|
6
|
|
|
11-06-2003 01:48 AM ET (US)
|
|
Diem brought up the similarity of this work to Flexible Sprites and I'd like to expand on this.
Both papers use the graphical models framework and both setup the problem relatively similarly. I think the essential difference between the two works can be summarized as follows. Jojic & Frey focus their efforts on powerful graphical model techniques to solve the problem in a 'brute force' method -- one that incorporates as little additional information about the problem and domain. Zhou and Tao explicitly incorporate knowledge of the domain and use graphical models only as a framework, designing an algorithm adapted to the domain.
As a specific example, consider that Jojic & Frey don’t even have a motion model – that is you could reorder all the frames of a video and their technique still works. On the one hand this means that you don’t need a motion model to do tracking (at least on simple cases) – so possibly on very deviant motion a motionless model would do better. On the other hand that information is available and easy to incorporate – so doing so, as Zhou and Tao did, makes their technique more robust in typical tracking problems.
I think Zhou and Tao got it right. The paper by Jojic & Frey is more interesting from a graphical models standpoint, but it kind of misses the point – rather then throw extremely powerful graphical models machinery at the problem why not analyze the domain more carefully? Anyway, my guess is that the work by Zhou and Tao is much more robust.
|
| Piotr
|
5
|
|
|
11-06-2003 01:32 AM ET (US)
|
|
Matt - I think that in the paper the definition of foreground is anything that moves independently of background (where the background is defined to be all objects that share the predominant motion). In an application setting you could use other knowledge to filter which foreground objects you think are significant.
|
| Matt Clothier
|
4
|
|
|
11-05-2003 10:55 PM ET (US)
|
|
I must say I am pretty impressed with the results. Certainly pinpointing the person in the middle frame of figure 10 is a great success (I totally agree with Sunny - I can't hardly see the person!). Anyway, ever since we have been presented with the idea of "sprites" since earlier this quarter, my interest has grown in this area. I think that by segmenting out certain part of an image and building them up as layers is a step in the right direction to classifying objects and motion.
One interesting things that I noticed was that they have made all background layers share the same motion. In most cases, I can see that this should work. However, I can imagine a scenario where a background object moves independent of the other background objects. Let's say that the camera is focused on a harbor watching boats leave. The primary application would probably be to watch when certain boats come and go. However, what if there is a buoy in the water that is bobbing up and down that should be part of the background? I guess the buoy could be made part of the foreground but then you would be tracking an undesirable object. It seems that in future work they could adapt a model in which the background layers are a bit "deformable" meaning that each layer would have a little freedom to move. However, such a model would have to be careful not to eliminate objects that are moving about as slow as the deformation would allow. Anyway, it is an interesting problem and I look forward to what future papers have to say regarding this.
|
| Diem Vu
|
3
|
|
|
11-05-2003 10:54 PM ET (US)
|
|
This paper looks like a variant of 'flexible sprite' with the ability to add, merge or reorder layers at run time. Multiple background layters may help to reduce computational cost. However, it is still unable handle non-rigid transformation.
|
| Sunny Chow
|
2
|
|
|
11-05-2003 09:51 PM ET (US)
|
|
While the paper does make a good number of assumptions, I am amazed at how it was able to continue tracking a person in the center image of Figure 10. The person has all but disappeared in that image.
The paper also makes the observation that background shapes cannot be deteremined without actual movement of foreground objects. While it may be true in the context of the paper, there does exist some information within a static background that can cue us on the "approximate" shape of the backgrounds (such as textures). I wonder how that extra information might be used to improve on this paper's methods.
|
| Jing Shiau
|
1
|
|
|
11-05-2003 08:34 PM ET (US)
|
|
I don't see why Figure 4 is such a hard problem. The object is only moving horizontally, and the background doesn't change. How would shadow affect tracking? The shadow didn't change in the synthetic video, and it's not blocking the object, so for all its worth, it's just a constant background (like trees in the other video sequences). I don't see reflection either... Also, why would transparency cause a big problem? If anything, wouldn't it aid tracking because a little bit of the object is still visible (appearence model should be able to pick it up)?
In the human tracking sequence, one person is wearing white and the other is wearing black. This should help the appearence model significantly. Would the algorithm be successful in differentiating the two persons if they are wearing the same color clothing?
Having said that, this is still a cool algorithm. Like the authors said, it models more possible interaction between foreground and background objects.
|
