Objectness to improve the bag of visual words model

The objectness measure identifies the image region that potentially contains any kind of object. This work studies several ways of using the objectness measure to improve the bag of visual words model for object recognition. The first approach is to extract descriptors only from the image regions with a high objectness score in order to obtain the vocabulary of visual words. The second approach is to weight the visual words from a standard vocabulary, according to the objectness measure computed on each image. The two approaches are also combined together using multiple kernel learning. Object recognition experiments are conducted to assess the performance level gained by integrating the objectness measure in the bag of visual words model. The empirical results show that the objectness measure can indeed improve the performance, by demoting the visual words located on the background of the image.