Adaptive VA model

The general model of visual attention (VA) is well suited for its universal detection behaviour in any environment but inferior to dedicated feature detectors in more specific environments. The goal here is to remedy this disadvantage by providing an adaptive visual attention model that, after its automatic tuning to a given environment during a learning phase, performs similarly well as a dedicated feature detector.

Methodology

The basic idea consists in providing the model with adaptation capabilities and then teaching this new adaptive VA system to produce the expected spots of attention. Two fundamental different approaches are supervised learning and unsupervised learning. With supervised learning, the adaptation parameters A are changed until expected spots of attention are detected. With unsuperwised learning, similar convergence can be expected by reinforcement learning.

a) football sequence

b) indoor sequence

c) traffic sequence

Figure 1: Adaptive VA model with the set A of parameters that must be tuned for good results.

Figure 2: Example frames of three video sequences and the related 6 spots of attention; the cheese diagram indicates the spot percentile contribution to saliency of intensity (black), chromaticity (red) and orientation (grey)

Results

The exploratory VA experiments performed with the generic VA model show that in presence of video sequences of very different nature, the maps provided differ enough and justify model adaptation. Therefore, the proposed adaptive visual attention model represents a challenging frame for further investigations and expected improvements in adaptive visual attention.

Publication

[1] H. Hügli & A. Bur, "Adaptive visual attention model", Proc. Conf. Image and Vision Computing New Zealand, Hamilton, Dec. 2007