Are summary statistics enough? Evidence for the importance of shape in guiding visual search

Abstract

Peripheral vision outside the focus of attention may rely on summary statistics. We used a gaze-contingent paradigm to directly test this assumption by asking whether search performance differed between targets and statistically-matched visualizations of the same targets. Four-object search displays included one statistically-matched object that was replaced by an unaltered version of the object during the first eye movement. Targets were designated by previews, which were never altered. Two types of statistically-matched objects were tested: One that maintained global shape and one that did not. Differences in guidance were found between targets and statistically-matched objects when shape was not preserved, suggesting that they were not informationally equivalent. Responses were also slower after target fixation when shape was not preserved, suggesting an extrafoveal processing of the target that again used shape information. We conclude that summary statistics must include some global shape information to approximate the peripheral information used during search.

Keywords:

• Summary statistics
• Visual search guidance
• Gaze contingent
• Eye movements
• Extrafoveal processing
• Shape

We thank Thitapa Shinaprayoon, Harneet Sahni, and Samuel Levy for their help with data collection, and Ruth Rosenholtz, Krista Ehinger, Shuang Song, and all the members of the Stony Brook Eye Cog lab for invaluable discussions. This work was supported by NSF grants [IIS-1111047] and [IIS-1161876 to G.J.Z], and NIH Grant [R01-MH063748] to G.J.Z.

We thank Thitapa Shinaprayoon, Harneet Sahni, and Samuel Levy for their help with data collection, and Ruth Rosenholtz, Krista Ehinger, Shuang Song, and all the members of the Stony Brook Eye Cog lab for invaluable discussions. This work was supported by NSF grants [IIS-1111047] and [IIS-1161876 to G.J.Z], and NIH Grant [R01-MH063748] to G.J.Z.

Notes

¹ Methods of computing summary statistics differ in the degree to which they break the spatial relationships between features; this breakage is pronounced in the single pooling region version of the Texture Tiling Model but far less so in more recent methods that employ multiple pooling regions (Freeman & Simoncelli, 2011; Rosenholtz, Huang, & Ehinger, 2012). Moreover, to the extent that higher-order statistics are used to compute summary representations, this breakage will never be entirely complete. However, while it is true that most methods do preserve spatial statistics to varying degrees, it is also true that these methods must discard some spatial relationship information if they are to explain the core phenomena of interest; if spatial information was preserved completely, the swapping of nearby features believed to largely determine perception in the visual periphery should not occur.

² In a related analysis we asked whether differences in first object fixations were due to a speed–accuracy trade-off. However, the time to fixate the first object was not reliably different between conditions, (β ≥ −3.50, SE ≥ 4.51, t ≤ 0.06, p ≥ .43), indicating that our evidence for guidance was not reflecting a trade-off between speed and accuracy.