Semantic Expectation Effects on Object Detection: Using Figure Assignment to Elucidate Mechanisms

Abstract

Recent evidence suggesting that object detection is improved following valid rather than invalid labels implies that semantics influence object detection. It is not clear, however, whether the results index object detection or feature detection. Further, because control conditions were absent and labels and objects were repeated multiple times, the mechanisms are unknown. We assessed object detection via figure assignment, whereby objects are segmented from backgrounds. Masked bipartite displays depicting a portion of a mono-oriented object (a familiar configuration) on one side of a central border were shown once only for 90 or 100 ms. Familiar configuration is a figural prior. Accurate detection was indexed by reports of an object on the familiar configuration side of the border. Compared to control experiments without labels, valid labels improved accuracy and reduced response times (RTs) more for upright than inverted objects (Studies 1 and 2). Invalid labels denoting different superordinate-level objects (DSC; Study 1) or same superordinate-level objects (SSC; Study 2) reduced accuracy for upright displays only. Orientation dependency indicates that effects are mediated by activated object representations rather than features which are invariant over orientation. Following invalid SSC labels (Study 2), accurate detection RTs were longer than control for both orientations, implicating conflict between semantic representations that had to be resolved before object detection. These results demonstrate that object detection is not just affected by semantics, it entails semantics.

Keywords: figure assignment; object detection; semantic conflict; semantic network; semantics; superordinate-level category.

Figure 1

Sample Bipartite Displays. In all stimuli, a portion of a well-known object was sketched on one “critical” side of the central border; this critical region was equally often on the left/right, in black/white, and upright/inverted. In these samples, the portions of the well-known objects are sketched on the right side of the central border in black (upright portions of a woman and an umbrella are shown in (A,B), respectively; inverted versions are shown in (C,D), respectively). Displays were presented on a medium gray background.

Figure 2

Trial Structures for labels-present and labels-absent experiments. (A) Trial structure for labels-present experiments (Studies 1 and 2). Following fixation, a label was displayed for 250 ms. The label was either valid or invalid. Valid labels denoted the object sketched in the displays at a basic level (e.g., “umbrella,” as depicted above); Invalid DSC labels (Study 1) denoted an unrelated object in a different superordinate-level category (e.g., “squirrel”); and Invalid SSC labels (Study 2) denoted an unrelated object in the same superordinate-level category (e.g., “envelope”). After a 500 ms blank screen, the test display was shown for either 90 ms or 100 ms (these durations were tested in separate experiments) and was followed by a 200 ms mask. The test display shown above depicts a portion of an upright umbrella sketched on the right side of the central border in black. During the experiments, the portions of common objects sketched on the critical sides of the borders were shown equally often on the left/right, in black/white, and upright/inverted. Task: Report the side on which they perceive a figure. The last, blank, screen was shown until response or 4 sec (timeout). (B) Trial structure for labels-absent experiments (control). Following fixation, a blank screen was displayed for 100 ms. As in labels-present experiments, the test display was shown for either 90 ms or 100 ms (durations tested separately) and was followed by a 200 ms mask. During the experiments, the portions of common objects sketched on the critical sides of the borders were shown equally often on the left/right, in black/white, and upright/inverted. Task: Report the side on which they perceive a figure. The last, blank, screen was shown until response or 3 s (timeout).

Figure 3

Results for control/labels-absent experiments. (A) Object Detection Accuracy (N = 121) and (B) Detection RTs (N = 113). Error bars represent standard errors. *** indicates p < 0.001 and ns indicates non-significance.

Figure 4

Results for Study 1: Invalid DSC Labels. (A) Object Detection Accuracy (N = 112) and (B) Detection RTs (N = 105). Error bars represent standard errors. *** indicates p < 0.001 and ns indicates non-significance.

Figure 5

Results for comparisons to control. (A–D) Comparisons to Study 1 and (E–H) comparisons to Study 2. Object detection accuracy difference scores following valid (A,E) and invalid (B,F) labels. Detection RT differences scores following valid (C,G) and invalid (D,H) labels. White asterisks indicate main effects of condition (experimental vs. control). Brackets and black asterisks indicate two-way interactions between conditions (experimental vs. control) and orientation. Error bars represent pooled standard errors. *** indicates p < 0.001, ** indicates p < 0.03, and * indicates p < 0.05.

Figure 6

Results for Study 2: Invalid SSC Labels. (A) Object Detection Accuracy (N = 113) and (B) detection RTs (N = 103). Error bars represent standard errors. *** indicates p < 0.001 and ns indicates non-significance.