Difference: PhilippeBeaudoin (1 vs. 8)

• In-depth study of motion alignment
  • It seems like distance-based motion alignment is a difficult problem: a punch to the left and a punch to the right will almost necessarily not get aligned right. A good alignment would aligh the extremas of these punches, but this is the "worst case" for a distance-based technique.
  • Investigate how other techniques behave: aligning velocities, accelerations, blending-based alignment.
  • Blending-based alignment: find an alignment so that for any blending ratio, the blended motion introduces/destroy as few frequency components as possible.
  • Blending-based alignment can probably be casted entirely in the frequency domain, an interesting theoretical project.
  • Thoroughly study n-way alignment
  • Build parameterized blend spaces
  • Perform a user-study to evaluate the quality of the resulting motion
  • Study the effect of modifying the blending parameter during the motion
  • Evaluate the physical correctness of the various approach, or with an evolving blending parameter
• Motion synthesis in wavelet space

Random thoughts

• Definition of "action", "motion" and "state" in a dynamical-system vocabulary
  • An active dynamical system D is a triplet (S,M,A) that contains a set of states, motions and actions. (none of these needs to be finite)
  • A motion m:S->S is a function of the state space into itself. This means that, given any specific motion from M, a state s can only be mapped towards a single other state s'. In other words, under a fixed motion, trajectory lines cannot cross.
  • An action a:M->M is a function of the motion space into itself. In other words, if we notice trajectory line that crosses, it means an action has been applied.
  • A projection of D towards D' = (S',M',A') is a transformation T:S->S' of the state space such that S' is of lower dimensionality than S. Moreover, for any m in M there exist m' in M' such that m (s) = s' <==> m' (T(s)) = T(s'). Similarly for the actions.
  • In other words, a projection reduces the dimensionality of the state space, but does not introduce crossings in the trajectories for any motion.
  • A reduced dynamical system D is such that it cannot be projected in any way.

About Me

Pierre Poulin

More information can be found on my Home Page

Animation links

• CharacterAnimationHome
• CAProjectIdeas
• CharacterAnimationGroup

Teaching links

Various Project Ideas

Personal Interests (beside CG)

Technical Links

Personal Preferences (details in TWikiVariables)

Related topics

• Phone: (604) 827-3992
• Office: X653
• Lab: X660 (Imager lab)
• Comment: Post-Doc with Michiel van de Panne in the Imager Lab (until January 2010).

• TeachingTips

• Refresh and get better with all this Machine Learning stuff

• Details and state of the art in dynamic character simulation
  • What is meant, exactly, by a controller?

Summer Project Ideas

• Some very abstract ideas...
  • When compressing, study if similar motions have a similar optimal truncated wavelet coefficient distribution.
  • If so, then use this fact (together with PCA?) to reduce the search space for various kinematic motion synthesis.
• Throwing and catching
  • Acquire mocap data for a character catching an object (a ball?)
  • Parameterize on:
    • Catching location
    • Object speed and mass
  • Kinematically synthesize a novel motion given arbitrary parameters
• Efficient real-time exploration of a distant motion-capture database
  • Motions are transmitted across a slow channel (ie. the internet)
  • Simultaneously display a bunch of huge bunch of animations (on the same ground plane, with a wide perspective view)
  • Use wavelet-based progressive motion compression to obtain adaptive level-of-detail
    • Motion far from the viewer is less detailed
    • Automatic and nice degradation when transmission slows down
  • Use streaming transmission
  • Spatially regroup animations according to similarity
  • Automatixally extract a motion hierarchy based on similarity
    • Display a sample motion for a a complete group of similar motion that can later be "opened up"
• Compression related projects
  • Develop and study more involved spatial compressions
    • PCA across the DOFs with pre- or post-wavelet compression (useful for large database compression)
    • Use automatic motion segmentation (Barbic et al. 2004) and pre-canned linear spaces adapted to the extracted clip
    • Dictionnary-based methods for large database (problem: building an efficient dictionnary of motion clips)
  • Study the importance of various errors in human perception of skeletal animations
  • Apply compression to complex animation data structures
    • Motion Graphs
    • Specialized database for synthesis of a specific kind of motion
  • Constaint enforcement during compression
    • Constrain extremal poses and derivatives (for inclusion in a "snap-together" style of motion graph, or for cycling motions)
    • 2-person interaction (dancing...) where animations constrain each other.
  • Efficient motion blending using a wavelet basis
  • Develop and study prediction-correction schemes (similar to MPEG motion compensation)
  • Build and use statistical motion models to increase compression efficiency
    • The model could be for a complete skeleton or part of a skeleton
    • The model could be specific (extracted from a small bank of similar motions) or generic (extracted from a large motion database)
    • The model could be for really short motion segments or for complete clips
    • Better yet, it could be multi-resolution: long clips for wide basis functions and short clips for narrow basis functions.

• Comment: Visiting student of Michiel van de Panne in the Imager Lab for the summer (until september 1st 2006).

• HerbertChow (undergrad)

• VancouverRestaurants
• My blog about boardgames (in french)

• Name: Philippe Beaudoin
• Login Name: beaudoin
• Email: beaudoin@csDELETEthisTEXT.ubc.ca
• Phone: (514) 268-7827
• Department: Imager Lab
• Comment: Visiting student of Michiel van de Panne in the Imager lab for the summer (until september 1st 2006).

About Me

I'll be spending the summer (May to August 2006) in the Imager Lab as a visiting student of MichielVanDePanne.