Predicating Scenes Where Unknown Object May Appear Based On Object Similarity

Scene prediction is the process of predicting the possible scenes in which the object may appear based on the contents of the image. Because of its realistic meaning and the broad application prospect, the scene prediction has the very important research value. In this paper, starting from the similarity of objects, the possible scenes in which the unknown object may appear were predicted based on the correlation between objects and scenes. The main work is as follows:Based on the similarity between objects, a method to predict the scenes in which the unknown object may appear was proposed. Firstly, Object Bank was applied to detect objects in the image. Then the similarity of unknown object and known object was calculated. After that, through the statistical probability of known objects appear in the corresponding scenes and by using Bayes’theorem and Probability transfer principle, the possible scenes in which the unknown object may appear was predicted. Experiments show that the proposed method has high recognition rate.The modified information entropy to evaluate the object detectors was put forward. In order to more effectively use the object detectors in the Object Bank to detect objects in the image, the evaluation method based on modified information entropy was proposed. The detectors chosen by this method have stronger ability to build the similarity between objects. It is illustrated by experiment that the detectors chosen by proposed method are more effective than the ones chosen by general choose method.A method to predict the position of object in the scene image based on Gaussian mixture model was proposed. Based on the response value of object image to the every object detector in Object Bank, the location where the object may appear was analysis by using likelihood and Gaussian mixture model. Then coupled with the prior probability of known object in the scene image, the posterior probability of unknown object’s location in the scene was calculated. Experiments show that the algorithm can effectively predict the position of object in the scene.