Purchase this article with an account.
Michael Slugocki, Allison Sekuler, Patrick Bennett; Evaluating Shape Representations using Machine Learning Systems. Journal of Vision 2018;18(10):1314. doi: https://doi.org/10.1167/18.10.1314.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Significant progress has been made toward understanding and modeling how the visual system constructs intermediate shape representations from constituent contour elements (Loffler, 2008; 2015). However, less is known about how various shape-encoding schemes affect performance in higher level tasks, such as object recognition. The current study examined how different shape representations constrain performance in a shape classification task. Specifically, we measured how changing encoding schemes affected the classification performance of several machine learning systems on a set of 10 unique shape classes, each consisting of 10,000 samples. The samples were constructed from a set of Radial Frequency (RF) contours, which allowed us to manipulate the low-level properties shared both within and between shape classes. We found that classification performance for sparse representations based upon the radial position of either positive or negative curvature extrema was generally high across machine learning methods, and was very robust to signal noise. In contrast, encoding schemes based upon angularity between neighbouring curvature extrema generally led to worse performance across all learning systems, and for a subset of these systems, introduction of noise into this representation type greatly affected performance. We currently are exploring how alternative encoding strategies can be used to learn to classify different families of shapes. These results highlight the utility in using machine learning methods to probe how different encoding schemas may contribute to the learning of shape identities.
Meeting abstract presented at VSS 2018
This PDF is available to Subscribers Only