Abstract
Real-world scenes are complex but lawful entities. From our experiences in the world, we know that blenders are more likely to be found in kitchens than forests, that cars are not generally found inside homes, that forks and knives are found near plates, etc. Research over the past 40 years has demonstrated that contextual associations influence object recognition accuracy, change the distribution of eye movements and modulate patterns of brain activations. However, the majority of these studies choose object-scene pairs from intuition because the statistical relationships between objects and scenes had yet to be systematically quantified. How do these intuitive estimations compare to actual object frequencies in the world? Ten observers were asked to list objects that were always, often, sometimes or never in each of 16 scene categories. These estimated object frequencies were then compared to the observed object frequencies in a fully labeled database of 3499 scenes (Greene, 2013). 62% of the objects listed by the observers corresponded to objects observed in the database. Although estimated frequencies were indeed correlated with estimated frequencies (r=0.63), participants systematically overestimated object frequency by 11% on average. Estimated object frequency was significantly greater than observed frequency for all scene categories (p<0.01). Furthermore, the estimation error did not systematically vary with object frequency (r=0.19), suggesting that this overestimation was not solely driven by infrequent objects. Altogether, these results speak to the richness of scene schemata and to the necessity of measuring object frequencies from scene databases because our intuitive estimations are unreliable.
Meeting abstract presented at VSS 2014