1 / 23

Activity Recognition Journal Club

Activity Recognition Journal Club. “Neural Mechanisms for the Recognition of Biological Movements” Martin Giese, Tomaso Poggio (Nature Neuroscience Review, 2003). Objective. Recognition of complex movements and actions using a neurophysiologically plausible and quantitative model

cameronk
Télécharger la présentation

Activity Recognition Journal Club

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Activity RecognitionJournal Club “Neural Mechanisms for the Recognition of Biological Movements” Martin Giese, Tomaso Poggio (Nature Neuroscience Review, 2003)

  2. Objective • Recognition of complex movements and actions using a neurophysiologically plausible and quantitative model • Biology has already generated a system that is robust and has high selectivity - let’s mimic it.

  3. Biological Intuitions • Separate dorsal and ventral streams http://upload.wikimedia.org/wikipedia/commons/thumb/c/c0/Gray722.png/300px-Gray722.png http://hebb.mit.edu/courses/8.515/lecture3/sld007.htm

  4. Biological Intuitions/Assumptions • Hierarchical architecture with increasing complexity along the hierarchy • Mainly feedforward • Activities and views are learned • Interaction between two streams not necessary for some recognition

  5. Biological Intuitions

  6. Artificial Neural Networks • Feedforward tr dv/dt = -v + F(W·u) • Recurrent tr dv/dt = -v + F(W·u + M·u)

  7. Form Pathway • Object Recognition Riesenhuber, Poggio 2002

  8. Form Pathway • Simple Cells • Modeled by Gabor Filters • Output via Linear threshold function • Complex Cells • ‘MAX’ function • Output via linear threshold function • Invariant bar detectors Giese, Poggio, AIM-2002-012, 2002.

  9. Form Pathway • View/Object Tuned Units • Radial Basis Function • u is the vector of the responses of the significant invariant bar detector • u0 signifies the preferred input pattern • C is a diagonal matrix with the elements Cu that are inversely proportional to the variance of the l-th component of u in the training set Giese, Poggio, AIM-2002-012, 2002.

  10. Form Pathway • Motion pattern neurons • Leaky integrator • ty dy/dt + y(t) = S f(un(t)) • Most active when input follows synaptically encoded sequence

  11. Motion Pathway • Represents dorsal stream • Similar hierarchy • Increasing complexity and invariance

  12. Motion Pathway • Lowest level • Optical flow, computed directly • Second Level • First Class • Translational flow (four tuned directions) • Second Class • Motion edges (horizontal and vertical) • MAX operator • Third Level • “Snapshot Neurons” • Modelled by RBF’s • Fourth Level • Motion Pattern Neurons

  13. Motion Pathway

  14. Model Parameters • Simple Cells • 8 orientations, 2 scales • s1=10, s2=7, k=0.35 • Motion Pattern & Motion Snapshot Neurons • ts=150ms Giese, Poggio, AIM-2002-012, 2002.

  15. Results

  16. Results

  17. Results

  18. Limitations • Does not address ‘attentional’ effects • No model for computing optical flow • Does not address interaction between form and motion streams • Form stream does not recognize point-light motion as per experimental data

  19. Recap

  20. Recap

  21. Example • Pattern (sn) • a • b • u(:,t+1) = sn(:,t) + w * tanh(u(:,t)); • v(:,t+1) = sum( tanh(u(1:3,t)) );

  22. Example

More Related