Is Whole Object Information Helpful for Scene Recognition?

Abstract

Scene recognition is one of the visual tasks, classifying a place category on an image. Scene images may contain various objects, and these objects tend to become clues to recognize the scene of the image. Therefore, many previous approaches for scene recognition use the object information that appeared in the image to improve the performance. Here, we raise a question of whether whole object information is helpful for scene recognition. To find the answer to the question, we conduct experiments on Places365, which is the largest scene recognition dataset consist of real-world images. To find the object classes which disturbed scene recognition, we utilize the Class Conversion Matrix, which is a deep learning approach. Finally, we found that some object classes may contribute to disturbing scene recognition. It indicates that not only making good use of object information, but also dropping disturbed object information is also important.