Statistical Models of Motion

Performance Animation from Low Dimensional Control Signals

Comments & Questions:

Using local models constructed from a database of human motion at run time to fill in the data for motion not captured by the markers sounds quite promising, but I need some further elaborations on what a heterogeneous database looks like, and why a local model works perfectly well with it while a global model doesn’t. -- Steven Chang

This was Michiel's AMoRe paper from November 2, 2005, so he may have a lot to say. The idea is a cute one, but again, since it is a database based approach, the quality of the captured animations depends on the heterogenousness of the database. Still though, applications could be realized for games with specific type of motion (e.g., a boxing game would only need a database of boxing type motions). Though not the major contribution of the paper, I like their Motion Analysis (4.1) section that describes the vision algorithm. One thing I wonder is that there seem to be a lot of parameters (alpha = 0.8 for the query metric, epsilon and Delta d for the k-nearest neighbour search, alpha and beta for the convex weights of the energy function). Is there any rhyme and reason for their choices of the values of these parameters? -- KenRose

A Data-Driven Approach to Quantifying Natural Human Motion

Comments & Questions:

The issues with the method given in the paper should be similar to the issues normally seen in Shape Recognition using backpropagation in neural networks. Therefore, it could be generally good for serving the purpose. However, the training data is of essence to make whole system work properly. From the result, it shows that SLDS works overall better than the rest, though I think the performances of different models are somehow determined by the complexities of the target motion, that is, motion for both training and testing; certainly, there is always room for improvments. -- Steven Chang

Reading this paper made me realize that I forgot a lot of statistics. I would definitely appreciate a small review in MoG, HMMs and SLDS. With respect to the paper, the idea seems interesting, but limited for large scale application since "the measures cannot be significantly better than the motion database of positive examples used to train them". This could be used for measuring the effectiveness of other animation systems though (e.g., rate the naturalness of Perlin noise). I like their idea of using a hierarchical group of features since that explicitly imbues semantics into the data that the models are classifying (that is, that this is human motion data and so there are associations between things like the tibia and femur and that this is not just a bunch of numbers). -- KenRose