Objects play a crucial role in our ability to categorize scenes. Two object properties are particularly important for predicting scene categorization accuracy: Object frequency and specificity for the given scene category. These object-occurrence statistics are often estimated from large-scale databases of labelled scene images. However, database measures do not always correlate with human ratings of specificity and frequency, opening up the question which of the two – databases or human ratings – better represent object-occurrence statistics in the real world. Here we therefore compared object frequency and specificity estimates gathered from human ratings (6-point Likert scale), scene image databases versus real-world environments (i.e., accurate object counts). Participants (N = 91) were asked to report the presence and number of 79 pre-selected objects in different rooms of their homes and rate their visibility and usage frequency. Intriguingly, while specificity measures were relatively consistent across participants’ self-reports and scene databases, we found significant differences between self-reports and databases with regard to object frequency. Especially smaller, non-stationary objects such as consumable goods (e.g., shampoo) and objects that are usually “hidden” or stored away (e.g., dishes or detergent) were underrepresented in scene image databases. Nonetheless, averaged human-rated object-occurrence statistics still outperformed both real-world and database statistics in predicting scene categorization accuracy, suggesting that scene categorization is based on prototypical object representations rather than real-world occurrence statistics. We therefore suggest considering the suitability of each source of object-occurrence statistics depending on the researchers’ goal and argue for more realistic and representative real-world scene image databases.