Alexander Muryy

Specular reflections and the estimation of shape from binocular disparity

Binocular stereopsis is a powerful visual depth cue. To exploit it, the brain matches features from the two eyes’ views and measures their interocular disparity. This works well for matte surfaces because disparities indicate true surface locations. However, specular (glossy) surfaces are problematic because highlights and reflections are displaced from the true surface in depth, leading to information that conflicts with other cues to 3D shape. Here, we address the question of how the visual system identifies the disparity information created by specular reflections. One possibility is that the brain uses monocular cues to identify that a surface is specular and modifies its interpretation of the disparities accordingly. However, by characterizing the behavior of specular disparities we show that the disparity signals themselves provide key information (“intrinsic markers”) that enable potentially misleading disparities to be identified and rejected. We presented participants with binocular views of specular objects and asked them to report perceived depths by adjusting probe dots. For simple surfaces—which do not exhibit intrinsic indicators that the disparities are “wrong”—participants incorrectly treat disparities at face value, leading to erroneous judgments. When surfaces are more complex we find the visual system also errs where the signals are reliable, but rejects and interpolates across areas with large vertical disparities and horizontal disparity gradients. This suggests a general mechanism in which the visual system assesses the origin and utility of sensory signals based on intrinsic markers of their reliability.

Key characteristics of specular stereo

Because specular reflection is view-dependent, shiny surfaces behave radically differently from matte, textured surfaces when viewed with two eyes. As a result, specular reflections pose substantial problems for binocular stereopsis. Here we use a combination of computer graphics and geometrical analysis to characterize the key respects in which specular stereo differs from standard stereo, to identify how and why the human visual system fails to reconstruct depths correctly from specular reflections. We describe rendering of stereoscopic images of specular surfaces in which the disparity information can be varied parametrically and independently of monocular appearance. Using the generated surfaces and images, we explain how stereo correspondence can be established with known and unknown surface geometry. We show that even with known geometry, stereo matching for specular surfaces is nontrivial because points in one eye may have zero, one, or multiple matches in the other eye. Matching features typically yield skew (nonintersecting) rays, leading to substantial ortho-epipolar components to the disparities, which makes deriving depth values from matches nontrivial. We suggest that the human visual system may base its depth estimates solely on the epipolar components of disparities while treating the ortho-epipolar components as a measure of the underlying reliability of the disparity signals. Reconstructing virtual surfaces according to these principles reveals that they are piece-wise smooth with very large discontinuities close to inflection points on the physical surface. Together, these distinctive characteristics lead to cues that the visual system could use to diagnose specular reflections from binocular information.

The Southampton-York Natural Scenes (SYNS) dataset: statistics of surface attitude

Recovering 3D scenes from 2D images is an under-constrained task; optimal estimation depends upon knowledge of the underlying scene statistics. Here we introduce the Southampton-York Natural Scenes dataset (SYNS: https://syns.soton.ac.uk), which provides comprehensive scene statistics useful for understanding biological vision and for improving machine vision systems. In order to capture the diversity of environments that humans encounter, scenes were surveyed at random locations within 25 indoor and outdoor categories. Each survey includes (i) spherical LiDAR range data (ii) high-dynamic range spherical imagery and (iii) a panorama of stereo image pairs. We envisage many uses for the dataset and present one example: an analysis of surface attitude statistics, conditioned on scene category and viewing elevation. Surface normals were estimated using a novel adaptive scale selection algorithm. Across categories, surface attitude below the horizon is dominated by the ground plane (0° tilt). Near the horizon, probability density is elevated at 90°/270° tilt due to vertical surfaces (trees, walls). Above the horizon, probability density is elevated near 0° slant due to overhead structure such as ceilings and leaf canopies. These structural regularities represent potentially useful prior assumptions for human and machine observers, and may predict human biases in perceived surface attitude.

'Proto-rivalry': how the binocular brain identifies gloss.

Visually identifying glossy surfaces can be crucial for survival (e.g. ice patches on a road), yet estimating gloss is computationally challenging for both human and machine vision. Here, we demonstrate that human gloss perception exploits some surprisingly simple binocular fusion signals, which are likely available early in the visual cortex. In particular, we show that the unusual disparity gradients and vertical offsets produced by reflections create distinctive ‘proto-rivalrous’ (barely fusible) image regions that are a critical indicator of gloss. We find that manipulating the gradients and vertical components of binocular disparities yields predictable changes in material appearance. Removing or occluding proto-rivalrous signals makes surfaces look matte, while artificially adding such signals to images makes them appear glossy. This suggests that the human visual system has internalized the idiosyncratic binocular fusion characteristics of glossy surfaces, providing a straightforward means of estimating surface attributes using low-level image signals.

Differential processing of binocular and monocular gloss cues in human visual cortex

The visual impression of an objects surface reflectance (“gloss”) relies on a range of visual cues, both monocular and binocular. Whereas previous imaging work has identified processing within ventral visual areas as important for monocular cues, little is known about cortical areas involved in processing binocular cues. Here, we used human functional MRI (fMRI) to test for brain areas selectively involved in the processing of binocular cues. We manipulated stereoscopic information to create four conditions that differed in their disparity structure and in the impression of surface gloss that they evoked. We performed multivoxel pattern analysis to find areas whose fMRI responses allow classes of stimuli to be distinguished based on their depth structure vs. material appearance. We show that higher dorsal areas play a role in processing binocular gloss information, in addition to known ventral areas involved in material processing, with ventral area lateral occipital responding to both object shape and surface material properties. Moreover, we tested for similarities between the representation of gloss from binocular cues and monocular cues. Specifically, we tested for transfer in the decoding performance of an algorithm trained on glossy vs. matte objects defined by either binocular or by monocular cues. We found transfer effects from monocular to binocular cues in dorsal visual area V3B/kinetic occipital (KO), suggesting a shared representation of the two cues in this area. These results indicate the involvement of mid- to high-level visual circuitry in the estimation of surface material properties, with V3B/KO potentially playing a role in integrating monocular and binocular cues.

The Southampton York natural scenes (SYNS) dataset

Humans are adept at estimating 3D scene geometry from a stereo image pair, or even from a single image. Computer vision algorithms are less good. Gaining traction on this problem requires a dataset that contains good quality images and ground truth data, and represents the complex and diverse scenes that we encounter. To this end we have developed the Southampton York Natural Scenes (SYNS) public dataset: https://syns.soton.ac.uk.

Brain processing of gloss information with 2D and 3D depth cues

Surface gloss information conveyed by image cues (i.e., highlights) has been shown to be processed in ventral and dorsal areas. In this study we used fMRI to distinguish the brain areas that selectively process 2D and 3D cues about surface gloss. We performed one experiment using 2D images of random objects with glossy surfaces where diffuse highlights could be presented rotated by 45 degree to make the object look matte. We also performed a second experiment with binocular cues where the specular reflections of the environmental on random shapes could have disparities coincident with the surface so to appear painted and thus making the object look matte. The same twelve participants took part in the two experiments where fMRI activations were measured over the whole brain with an Echo-Planar Imaging sequence (32 slices, TR 2000 ms, TE 35 ms, voxel size 2.5 × 2.5 × 3 mm). We performed Multi-Voxel Pattern Analysis to test whether a classifier trained to discriminate glossy vs. matte objects with 2D cues can still discriminate with 3D cues, and vice versa. We found transfer effects from 2D to 3D cues in early (V1, V2) and dorsal visual areas (V3d, V3A/B, V7, IPS). This transfer suggests the presence of circuits processing gloss independently of the type of cues in dorsal areas only. We did not find transfer from training with 3D cues to 2D cues, suggesting that stereoscopic information related to gloss has a pattern of activation that is additional to the representation of gloss. Meeting abstract presented at VSS 2015.