CORRECTIONS TO: Pramod, R. T., & Arun, S. P. (2016). Object attributes combine additively in visual search.
Journal of Vision, 16(5):8, 1–29, doi:
10.1167/16.5.8.
The main finding of our study is that the dissimilarity between multipart objects is a linear sum of pair-wise dissimilarities between their parts. Across several experiments, we reported that the dissimilarity between symmetric objects is larger by a fixed offset from that predicted by the part summation model. While devising follow-up studies, we discovered an interpretational error that invalidates our conclusions regarding symmetric objects. However, all other results remain essentially unchanged.
This error arose from the manner in which we were accounting for part relations in our model. According to our model, the net dissimilarity between objects
AB and
CD made from parts (
A,
B) and (
C,
D) respectively is given by the equation:
Here, c(.,.) is the dissimilarity between parts at corresponding locations, x(.,.) is dissimilarity between parts at opposite locations, and w(.,.) is the dissimilarity between parts within an object. In our original manuscript we had considered the contribution of a given pair of parts to be 1 if that part pair was present, and 0 if it was absent.
Specifically, even when part pair occurred twice (such as with symmetric or mirror objects), we still kept its contribution as 1. However, we realized that this is incorrect: The contribution of the repeated part pair should be taken twice since it contributes two times to the net dissimilarity. Thus, for the revised model, the net dissimilarity between a pair of symmetric objects (e.g., AA and CC) is given by: d(AA,CC) = 2c(A,C) + 2x(A,C) + w(A,A) + w(C,C) + constant, whereas in the original model the terms for c(A,C) and x(A,C) would be scaled by 1 and not 2.
On fixing this error and fitting this revised model to the data across all experiments, we found that the part summation model also explains the dissimilarities between symmetric object pairs without affecting its quality of fit on the other object pairs, as can be seen in the figures below. This improves the model performance overall as well. We note in addition that both in the original and updated results, the perceptual dissimilarity between mirror object pairs is consistently underestimated by the model. Thus, it appears that mirror confusion cannot be fully explained using the part summation model regardless of this error (i.e., whether part pairs are counted once or twice).
To sum up, we had originally concluded that symmetry increases distinctiveness by a fixed amount. However, in light of the revised model, this conclusion is no longer correct. Rather, our results show that dissimilarities between symmetric objects can be explained entirely using part summation.
Citation: Pramod, R. T., & Arun, S. P. (2016). Corrections to: Object attributes combine additively in visual search.
Journal of Vision, 16(11):21, 1–4, doi:
10.1167/16.5.8.