English
 
User Manual Privacy Policy Disclaimer Contact us
  Advanced SearchBrowse

Item

ITEM ACTIONSEXPORT

Released

Poster

A model for multi-stable dynamics in action recognition modulated by integration of silhouette and shading cues

MPS-Authors
There are no MPG-Authors available
External Ressource

Link
(Any fulltext)

Fulltext (public)
There are no public fulltexts stored in PuRe
Supplementary Material (public)
There is no public supplementary material available
Citation

Fedorov, L., & Giese, M. (2016). A model for multi-stable dynamics in action recognition modulated by integration of silhouette and shading cues. Poster presented at Twenty-Fifth Annual Computational Neuroscience Meeting (CNS*2016), Jeju, South Korea.


Cite as: http://hdl.handle.net/21.11116/0000-0000-7B46-8
Abstract
The visual perception of body motion can show interesting multi-stability. For example, a walking body silhouette (bottom inset Fig. 83A) is seen alternately as walking in two different directions [1]. For stimuli with minimal texture information, such as shading, this multi-stability disappears. Existing neural models for body motion perception [2–4] do not reproduce perceptual switching. Extending the model [2], we developed a neurodynamic model that accounts for this multi-stability (Fig. 83A). The core of the model is a two-dimensional neural field that consists of recurrently coupled neurons with selectivity for instantaneous body postures (‘snapshots’). The dimensions of the field encode the keyframe number θ and the view of the walker ϕ. The lateral connectivity of the field stabilizes two competing traveling pulse solutions that encode the perceived temporally changing action patterns (walking in the directions ±45°). The input activity of the field is generated by two visual pathways that recognize body postures from gray-level input movies. One pathway (‘silhouette pathway’) was adapted from [2] and recognizes shapes, mainly based on the contrast edges between the moving figure and the background. The second pathway is specialized for the analysis of luminance gradients inside the moving figure. Both pathways are hierarchical (deep) architectures, built from detectors that reproduce known properties of cortical neurons. Higher levels of the hierarchies extract more complex features with higher degree of position/scale invariance. The field activity is read out by two Motion Pattern (MP) neurons, which encode the two possible perceived walking directions. Testing the model with an unshaded silhouette stimulus, it produces randomly switching percepts that alternate between the walking directions (±45°) (Fig. 83B, C). Addition of shading cues disambiguates the percept and removes the bistability (Fig. 83D). The developed architecture accounts for the disambiguation by shape-from shading.