Measuring Observer Response to Object-Scene Disparity in Composites

Image composites are combinations of image elements from different sources, often combined in a manner to give the appearance of a single, coherent image. This assesses the impact of low-level image feature offsets on observer response with respect to realism of image composites. The response to selected features, namely exposure, contrast and Correlated Colour Temperature (CCT), is evaluated in a series of 3 experiments, each employing 25 human observers. A total of 10890 data points are analysed, 3630 for each experiment, and psychometric functions are fit to this data in order to parametrise and quantify the relationship between human responses and the amount of disparity between object and scene. Average thresholds and their confidence intervals for each of the image features are then presented and discussed, notably indicating a degree of observer variance in realism responses, particularly in the presence of negative contrast disparities. Exposure, as well as CCT offsets are found to be more readily detected, the latter also contributing to some false positives at high offsets, due to illuminationreflectance ambiguity. The resulting thresholds and confidence intervals can be utilised in creating realistic composites, as well as understanding the impact of different features on observers' perception of composite realism.