Abstract
Body movements are characterized by specific sequences of complex optic flow patterns. Computational models for the perception of static shapes have demonstrated that recognition performance can be significantly improved by choosing an appropriate dictionary of mid-level shape-components (see abstract by Serre Poggio, 2005). Preliminary results suggest that such shape-tuned units are consistent with recent physiological data collected in V4 (see abstract by Cadieu et al, 2005). We test if the visual recognition of complex body movements from optic flow might also benefit from optimized motion-component units.
METHOD: We employ a physiologically inspired learning algorithm for the optimization of mid-level motion detectors of a hierarchical model for the recognition of human actions (Giese Poggio, 2003). In the proposed algorithm, competing units are associated with a memory trace that reflects their recent synaptic activity. The model is presented with movies showing a human action (i.e. walking): the trace from units that are behaviorally-relevant is increased while the trace from the others is decreased. Units whose memory trace falls below a critical threshold are randomly replaced.
RESULTS: When presented with movies showing human actions, the model generates a dictionary of mid-level motion-component units that lead to a significant improvement of the recognition performance. For the special case of walking, many of the units' preferred stimuli were characterized by horizontal opponent motion, consistent with a recent experimental study showing that opponent horizontal motion is a critical feature for the recognition of these stimuli (Casile Giese, 2003).
CONCLUSION: Like for the categorization of static shapes, recognition performance for human actions is improved by choosing optimized mid-level motion features. In addition, the extracted features might predict receptive field properties of complex motion-selective neurons, e.g. in areas MT and MSTl.
