The role of global cues in the perceptual grouping of natural shapes

Abstract

Perceptual grouping of the bounding contours of objects is a crucial step in visual scene understanding and object recognition. The standard perceptual model for this task, supported by a convergence of physiological and psychophysical evidence, is based upon an association field that governs local grouping, and a Markov or transitivity assumption that allows global contours to be inferred solely from these local cues. However, computational studies suggest that these local cues may not be sufficient for reliable identification of object boundaries in natural scenes. Here we employ a novel psychophysical method to assess the potential role of more global factors in the perceptual grouping of natural object contours. Observers were asked to detect briefly presented fragmented target contours in oriented element noise. We employed natural animal shape stimuli, which in addition to local grouping cues possess global regularities that could potentially be exploited to guide grouping and thereby improve target detection performance. To isolate the role of these global regularities we contrasted performance with open and closed control target stimuli we call local metamers, as they afford the same local grouping cues as animal shapes. We found that performance for closed metamers exceeded performance for open metamers, while performance for animal targets exceeded both, indicating that global grouping cues represented in higher visual areas codetermine the association between orientation signals coded in early visual cortex. These results demand a revision to the standard model for perceptual grouping of contours to accommodate feedback from higher visual areas coding global shape properties.

Shape from Contour: Computation and Representation

“…Recently, Elder et al (2018) introduced two innovations in the psychophysical method of Field et al (1993) to address these issues. First, instead of employing simple geometric stimuli such as circles or Ss, they employed fragmented outlines of animal shapes, natural stimuli to which we know the human visual system is acutely tuned (see Section 1).…”

“…One limitation of backward-masking and TMS studies is that the transient onset of a strong visual mask or the repeated application of a focused magnetic field may generally disrupt processing and reduce performance on a range of tasks by introducing noise or distracting attention, and thus an observed deficit does not demonstrate that the feedback process is specific to the visual task under study (shape processing, in our case). To address this issue, Drewes et al (2016) developed a novel repetition methodology based on the animal and metamer stimuli of Elder et al (2018). In this method, a fragmented shape stimulus is displayed briefly in random dynamic noise, and the observer's task is to distinguish whether the shape is an animal or a metamer.…”

Learning compact generalizable neural representations supporting perceptual grouping

“…Here, we show that standard DCNs struggle to match the superior performance, sample efficiency and generalizability of the proposed recurrent V1Net-1L model on 2 contour integration tasks. On this note we believe that the utility of bio-inspired recurrent models such as V1Net should be explored in overcoming some of the recent deficiencies (where texture is biased over shape) observed in DCNs [21,6,4,45,40] by emphasizing extended contour processing through recurrent horizontal interactions [44,14,15].…”

“…V1Net models in our experiments were trained with 5 recurrent iterations. The kernel width of excitatory connections W exc is [15,15], which is thrice the size of our input kernel W xh in order to support far long-range excitatory connections. Inhibitory and divisive gain control connections are implemented by [7,7] separable convolution kernels W inh and W div respectively.…”

Radial frequency patterns describe a small and perceptually distinct subset of all possible planar shapes

“…We used population Monte Carlo (Cappé, Guilin, Marin, & Robert, 2004), starting with 391 circles. At each step of the algorithm, each of the 391 outlines was randomly perturbed using the technique developed by (Fruend & Elder, 2013, see also Elder, Oleskiw, & Fruend, 2018) and the resulting, perturbed outlines were re-sampled based on the fit with the target distribution. This procedure leads to a fair sample from the maximum entropy distribution of closed, non-intersecting shapes that match the marginal turning angle distribution of the animal outlines.…”