An experimental evaluation of visual similarity for HDR images

In this paper, we investigate visual similarity for high dynamic range (HDR) images. We collect crowdsourcing data through a web-based experimental interface, in which the participants are asked to choose one of the two candidate images as being more similar to the query image. Triplets forming the query-and-candidates sets are obtained by random sampling from existing HDR data sets. Experimental control factors include choice of tone mapping operator (TMO), choice of distance metric, and choice of image feature. The image features that we experiment with are chosen from the features that are commonly used in the usual low dynamic range setting including features learned via Convolutional Neural Networks. The set of image features also includes combined features where the combination coefficients are estimated using logistic regression. We compute correlations between human judgments and quantitative features to understand how much each feature contributes to visual similarity. Combined features yield nearly 84% agreement with human judgments when applied on tone mapped images. Though we observed that using common features directly on raw or linearly scaled HDR images yield subpar correlation estimates compared to using them on tone mapped HDR images, we did not observe significant effect due to the choice of TMO on the estimates. As an application, we propose an improvement to style-based tone mapping for more correctly imparting desired styles to HDR images with different characteristics.