Image Appeal Revisited: Analysis, New Dataset, and Prediction Models

There are more and more photographic images uploaded to social media platforms such as Instagram, Flickr, or Facebook on a daily basis. At the same time, attention and consumption for such images is high, with image views and liking as one of the success factors for users and driving forces for social media algorithms. Here, "liking" can be assumed to be driven by image appeal and further factors such as who is posting the images and what they may show and reveal about the posting person. It is therefore of high research interest to evaluate the appeal of such images in the context of social media platforms. Such an appeal evaluation may help to improve image quality or could be used as an additional filter criterion to select good images. To analyze image appeal, various datasets have been established over the past years. However, not all datasets contain high-resolution images, are up to date, or include additional data, such as meta-data or social-media-type data such as likes and views. We created our own dataset "AVT-ImageAppeal-Dataset", which includes images from different photo-sharing platforms. The dataset also includes a subset of other state-of-the-art datasets and is extended by social-media-type data, meta-data, and additional images. In this paper, we describe the dataset and a series of laboratory- and crowd-tests we conducted to evaluate image appeal. These tests indicate that there is only a small influence when likes and views are included in the presentation of the images in comparison to when these are not shown, and also the appeal ratings are only a little correlated to likes and views. Furthermore, it is shown that lab and crowd tests are highly similar considering the collected appeal ratings. In addition to the dataset, we also describe various machine learning models for the prediction of image appeal, using only the photo itself as input. The models have a similar or slightly better performance than state-of-the-art models. The evaluation indicates that there is still an improvement in image appeal prediction and furthermore, other aspects, such as the presentation context could be evaluated.