A strongly coupled architecture for contextual object and scene identification

The context-centered approach to object detection and recognition is based on the intuition that the contextual information of real-world scenes provides relevant information for these tasks. This intuition is supported by psychophysical experiments in human scene perception and visual search, which provide evidence that the human visual system uses the relationship between the environment and the objects to facilitate object recognition. Here, we use a probabilistic model to investigate the possible interactions between object class hypotheses and scene class hypotheses in a visual system. The architecture of the model is based on separate modules interacting with each other via feedforward and feedback connections. A competitive-priors structure is used to implement the feedback connections.