My understanding was that the bidirectional search makes the motion synthesis fast enough that it’s nearly real time, so the sketch interface is just drawing spacetime curves as per normal and then running a motion through them. That’s what it looks like is going on on the video, in any event.

One thing that I don’t understand is exactly how the sketch interface is implemented, I understand what it does, just not how it uses the bidirectional search to find it.

1. Laplacian operator is a low pass filter, so any close curve will soon turn into a circle and open curve into a straight line ? ( by Eigenvalues analysis ) and probably too soon. So it is that our actors will soon be forced into moving in circle or straight line ?

2. What exactly are the soft-constraints in the solver ? does it mean that some points are allowed to move at half the allowable speed ?

3. The whole beauty of the “Laplace Solver ” is messed up in equation 4, it is no more positive definate or probably diagonal dominant. I don’t understand why an elegant sparse and NxN matrix was converted into an overdetermined system of equation and Pseudo-Inverse came into scene. I am not sure, if purist will still call it Laplacian Editor.

4. The final dampener was that the reported work was done in 2D only. Equation 2 used only (x,y) coordinates and that also in local coordinate system. Never heard someone doing global operation with local system. I am thoroughly confused with their formulation. Something is surely missing in the paper.

It is well known fact that any generalization from lower dimension to higher dimension is non-trivial in most cases. I surely will not buy the arguments of the
authors in section 7, para #2.

5. Many strong sentences have been used in the paper without any references. “Motion path is notorious for its exponential computational complexity” should have been substantiated with citations and better explanation.

After re-reading the geostatistics paper, I think I can sum up the intuition for kriging in relatively few words. The overall goal is to use least-squares regression to minimize the variance of the prediction error. Universal kriging assumes a linear relationship for the spatial dependence of values in the random field (i.e. it linearly interpolates the sampled values). Using this expectance of spatial dependence combined with a variogram (a dissimilarity function of distance, estimated with the samples), we can form a geometrical probability model that will have minimized predictive error variance from kriging. With this statistical model, an optimal blending kernel can be computed that best compensates for the similarity (rather than just fitting a kernel to the nearby samples). This model is intrinsically stationary, or invariant, so it can be maintained as a character moves around.

Although I’m certainly not ready to implement this yet, I feel like I have a much better grasp on the ideas. Kriging threw so many new terms at me that it seemed more difficult than it really is; it really just comes down to a glorified form of least squares.

I tried to explore some of the issues that we had discussed in the class on Monday. One of the thing was “Laplacian Editing” which now I understand is the extremely simple and may not take more than 100 lines of C++ code to implement.

Here is the video from one of the authors which will explain what it does:
http://visgraph.cse.ust.hk/projects/dual_laplacian

There are some issues which authors are hiding or not forthcoming is that “Volume preservation ” issue. but I thing there are always workaround or engineering solution (for example distort the model from the invisible side and preserve the shape from visible side )

I think the authors have confused the use of ‘terminate’ and ‘converge’. The penultimate paragraph in that column says –
“Hence, one tree may surpass the cut and terminate much earlier than the other.” and ” It is also possible that one search converges more quickly to the optimal solution than the other”
What I interpret from that is
– termination is when one tree hits the cut -> then just shift the cut towards the other side
-if both converge before then maybe lookup the merge tables and find if there are any merge-able nodes

This paper presented a relatively simple, intuitive framework for simultaneously editing motions of multiple characters, while enforcing constraints using a Laplacian editing method. Although most of the techniques used in this paper are not entirely novel, I still found it to be a great paper because of the authors’ attention to details I’d not previous thought of; not to mention the incredibly simple yet full featured user interface.

Some of the things they paid attention to in implementing their framework that stood out: paying attention to possible flipped tangents, what to do when a user tries to edit stationary movement, incremental updates instead of searching well connected motion graphs, automatically inserting extra loops of the walking motion to retain naturalness if a sequence is dragged out too long. Also, their method of editing motion addresses something that I thought was strange when reading the motion graph paper last week, specifically the part about not being able to synthesize sharp turns unless the data contains examples of a sharp turn; seem like they solved that problem.

Also, as far as questions go, the one overall question I want to know is why this works? Why do geospatial interpolation methods map to plausible motion?

To really understand this paper, you would need an in-depth understanding of statistical methods. The paper relies heavily on a comparison of kriging to radial basis kernels to support it’s claims of being ‘better’ than previous methods and also to provide a groundwork for understanding their paper.

One of the most interesting things about this work is their integration of previously proven methods from other fields into animation. At first glance, it appears to provide very quick and plausible motions and a real-time constraint solving mechanism that needs little, if any, post-processing correction.

I just LOVED the UI in both the bidirectional search and sync multi-character papers. Simple, sweet and intuitive! I think I’ll summarize these papers on wednesday instead. I tried looking at Geostatistical motion interpolation since it sounded cool, but terms like gaussian process regression, intrinsic stationarity, variogram and the statistics knowledge one needs to really understand it made me cry.

points worth discussing:
i) are per-frame IK techniques more “real-time” than techniques that have a more global view (such as spacetime)?
ii) importance of “fast synthesis” and being able to “edit at real time”; are these one and the same?
iii) can temporal and spatial constraints/requirements be traded for each other? what i’m thinking here is related to the slicing paper. one of the problems was that the “mass” could not be regulated(when an object was being carried). one way to enforce the importance of the mass is either making the steps smaller (spatial editing) or taking slower steps (temporal editing). has this been done before?