The role of spatial frequency in ultra-rapid scene categorization

Publisher Website
Google Scholar
Abstract
Humans are very good at deciding whether briefly presented natural scenes contain an animal. ERP recordings indicate that the underlying processing can be performed in around 150 ms [Thorpe, Fize & Marlot, Nature, 381, 520–522 (1996)] but the interpretation of ERP differences can be problematic [Johnson & Olshausen, J.Vision, 3, 499–512 (2003)]. However, we recently reported that behavioural reaction times can be even faster than the differential ERP signal if subjects are asked to move their eyes to the side where an animal had appeared. After a random fixation interval and a 200 ms gap period, two photographs were flashed for 20 ms on either side of fixation, followed by peripheral fixation marks indicating the possible saccade landing positions. Performance was remarkably good with the fastest reliable saccades occurring in only 130 ms after stimulus onset [Kirchner, Bacon & Thorpe, Perception, 32 suppl. 170, (2003)]. In the present experiments we used both this choice saccade task and go/nogo and choice reaction time paradigms with manual responses to investigate the role of different spatial frequency components in this form of rapid visual processing. We again found evidence that subjects can reliably initiate saccades towards the side with the animal in under 150 ms. Furthermore, by using low-pass, high-pass and band-pass filtered images we were able to show that while both mid and high spatial frequency components contribute to classification accuracy, low spatial frequencies alone resulted in near-chance performance. Such data rules out the possibility that animal detection could be based on very coarse information. Instead, it would seem that features involving greater detail are required.

This publication has 0 references indexed in Scilit: