Contribution of Low, Mid and High-Level Image Features of Indoor Scenes in Predicting Human Similarity Judgements

Human judgments can still be considered the gold standard in the assessment of image similarity, but they are too expensive and time-consuming to acquire. Even though most existing computational models make almost exclusive use of low-level information to evaluate the similarity between images, human similarity judgements are known to rely on both high-level semantic and low-level visual image information. The current study aims to evaluate the impact of different types of image features on predicting human similarity judgements. We investigated how low-level (colour differences), mid-level (spatial envelope) and high-level (distributional semantics) information predict within-category human judgements of 400 indoor scenes across 4 categories in a Four-Alternative Forced Choice task in which participants had to select the most distinctive scene among four scenes presented on the screen. Linear regression analysis showed that low-level (t = 4.14, p < 0.001), mid-level (t = 3.22, p< 0.01) and high-level (t = 2.07, p < 0.04) scene information significantly predicted the probability of a scene to be selected. Additionally, the SVM model that incorporates low-mid-high level properties had 56% accuracy in predicting human similarity judgments. Our results point out: 1) the importance of including mid and high-level image properties into computational models of similarity to better characterise the cognitive mechanisms underlying human judgements, and 2) the necessity of further research in understanding how human similarity judgements are done as there is a sizeable variability in our data that it is not accounted for by the metrics we investigated.