A Goal Oriented Attention Model for Efficient Object Search

Abstract

Prior knowledge of the target accelerates target detection in visual search tasks. This paper suggests a new computational model which biases the bottom-up features with known target representation so as to make the target more salient and to speed up object search. The proposed model consists of two of models, learning model and searching model. Learning model is incrementally learns and memorizes primitive features of target object and yields trained data, and searching model finds desired targets through biasing feature maps and saliency map for selectively attending to a target object. The information in trained data is used as a biasing signal. In order to evaluate the performance of our model, we compared our model with previous bottom-up model and trained model in top-down guided search. Average number of false detections before target found was used as a performance criteria in our experiments. The results show that our model successfully finds desired target in natural cluttered scenes faster than previous models.