Abstract
The perceptual decisions we make during the course of the day engage automatic mechanisms of scene and place recognition: navigating through the environment, searching for an object, etc. However, studies of scene understanding have focused on scene classification; a task that may not reflect what we do in our daily lives. What is the perceptual efficiency of such decisions and to what degree do low-level image features correlate with these judgments? We have selected 3 dimensions that reflect common judgments made through environmental interactions: the degree to which things can be hidden (camouflaged) in a scene, how well one can navigate and the scene temperature. 500 outdoor images were chosen from a database of 5000 images to reflect membership along these dimensions. Images were ranked along each dimension. Subject agreement in ranks was good and around 100 images per dimension were found to be prototypical (low variance at dimension extremes). There was no correlation of image ranks between dimensions, suggesting that the features used or their complexity vary. The prototype images were run in a yes/no task for each dimension with variable presentation times between 20 and 80ms followed by a mask. We found a 15% performance advantage for temperature dimension over the other dimensions at 20ms that vanished by 60ms, suggesting that image features responsible for this task (e.g. color) are available earlier than properties related to the others. Detecting a navigational route had the steepest performance gain (28%) for increasing presentation times, indicating that the relevant features become massively available between 20 and 40ms. Correctly evaluating potentiality for camouflage requires more exposure time as it had a shallower gain over presentation times, indicating an even later availability of pertinent features. We show the extent to which low-level image features are correlated with the degree of temperature, navigation and temperature in a scene.
Grateful thanks to Antonio Torralba, Molly Potter, Ruth Rosenholtz and Nancy Kaniwsher for fruitful discussions.