Advanced Multimedia Processing (AMP) Lab, Cornell University

Object-Driven Image Group Annotation



Takayuki Baba and Tsuhan Chen




We are developing a method for annotating image groups in real-world photo image databases. It is often difficult to categorize a single image into an appropriate scene category because of the limitation in the object cues observable from a single image. Fortunately, it often happens that several images are consecutively taken in a scene. We consider these images taken in the same scene as an image group. If objects recognized from other images in this image group could be used, it would become possible to categorize the image group into an appropriate scene category due to the increased object cues. We show an example in a figure. It is difficult to categorize the single image in Fig. (a) into “city” or “country” because only a car and a road can be seen. However, we can easily categorize the image group in Fig. (b) into “city” because several buildings can be seen in 1st photo and 3rd photo in the same group. Inspired by this motivation, our work focuses on annotating image groups instead of annotating a single image individually. We are developing a novel method for annotating image groups using all objects recognized from images in these groups.