Paper Reading #12

Comments:
Comment #1
Comment #2
Reference:
Title: Cosaliency: Where People Look When Comparing Images
Author: David E. Jacobs, Dan B. Goldman, Eli Shechtman
Venue: UIST 2010
Summary:
Photographic triage refers to saving or deleting photos based on each photo's priority, or taking another photo when needed. This is a common task for photographers. This paper presents a potential solution to aid in this task. The paper proposes a "learned model for calculating the importance, or saliency, of image pixels in the context of other images." The article classifies this feature as cosaliency. Most research in image saliency only considers one image at a time, while this paper examines two images next to each other. The paper then classifies the term "goal map" to signify the user-generated cosaliency maps because, intuitively, these are what the program strives to obtain. The model uses eight features to classify images: Gaussian Prior, Contrast, Faces, Oliva Saliency, and Judd Saliency for single images, and Flow Divergence, Nearest Neighbor Error, and Nearest Neighbor Incoherence for multiple images. The authors had 198 people participate in the study, and cosaliency outperformed saliency by a good portion. The limitations were that this triage may not be appropriate for all types of images (such as sporting images), it only measured the effectiveness of static detail crops, and it doesn't explore the non photorealistic visualization of cosalient images.

Discussion:
This paper was a bit more boring than the previous papers I've read recently. This seems like a somewhat useful feature, but it really doesn't strike home to me. I don't mind leaving images as they are, and never would need something to crop the images I do crop. Also, this seems like a bit much to write a whole paper about. It just doesn't seem that important or novel. The math and ideology just didn't seem overly complex. Overall, this paper is ok, but it doesn't excite me nor make me want to keep track of things like this.